Remote Sensing Big Data for Water Environment Monitoring: Current Status, Challenges, and Future Prospects

Abstract

Accurate water extraction and quantitative estimation of water quality are two key and challenging issues for remote sensing of water environment. Recent advances in remote sensing big data, cloud computing, and machine learning have promoted these two fields into a new era. This study reviews the operating framework and methods of remote sensing big data for water environment monitoring, with emphasis on water extraction and quantitative estimation of water quality. The following aspects were investigated in this study: (a) image data source and model evaluation metrics; (b) state-of-the-art methods for water extraction, including threshold-based methods, water indices, and machine learning-based methods; (c) state-of-the-art models for quantitative estimation of water quality, including empirical models, semi-empirical/semi-analytical models, and machine learning-based models; (d) some shortcomings and three challenges of current remote sensing big data for water environment monitoring, namely the new data gap caused by massive heterogeneous data, inefficient water environment monitoring due to “low spatiotemporal resolution,” and low accuracy of water quality estimation models resulting from complex water composition and insufficient atmospheric correction methods for water bodies; and (e) five recommendations to solve these challenges, namely, using cloud computing and emerging sensors/platforms to monitor water changes in intensive time series, establishing models based on ensemble machine learning algorithms, exploring quantitative estimation models of water quality that couple physics and causality, identifying the missing elements in water environment assessments, and developing new governance models to meet the widespread applications of remote sensing of water environment. This review can help provide a potential roadmap and information support for researchers, practitioners, and management departments in the theoretical exploration and innovative application of remote sensing big data for water environment monitoring.

Key Points

• Progress of remote sensing big data for water environment monitoring are reviewed

• The state-of-the-art methods of water extraction and water quality estimation are classified, evaluated and discussed

• We summarized three bottlenecks and five promising roadmaps of remote sensing big data for water environment monitoring

Plain Language Summary

The intensifying conflict between people and water use generates a great demand to monitor water volume and water quality, with the key scientific issues of water extraction and water quality estimation. Specialized technology, called remote sensing big data, has extracted water body and estimated water quality worldwide in a broad-scale, rapid, and economical way. In order to understand the current status, challenges, and future prospects of remote sensing big data for water environment monitoring, we review image data source and model evaluation metrics, state-of-the-art methods for water extraction and water quality estimation. We concluded that remote sensing big data have great potential to accurately extract water and estimate optically active water matters. However, the lack of remote sensing images caused by clouds and rainy weather, the low reflectance of water, the storage and processing of remote sensing big data, and the estimation of non-optically active water parameters are four major restricting factors. Data-driven methods based on remote sensing big data, machine learning and cloud computing provide the promising applications for intensive long-term water dynamics and water quality estimation. Further, physically driven methods based on radiative transfer equation and data assimilation provide potential solutions for the quantitative estimation of water quality. Future remote sensing big data efforts for the water environment should take these results into account.

1 Introduction

Various water bodies (e.g., oceans, rivers, lakes, reservoirs, and groundwater) are the sources of life, foundations of ecology, and keys to production (Central Committee, State Council, CPC, 2011). The water environment plays a vital role in the sustainability of both human and ecological systems (Huang et al., 2018). With the continuous increase in the global population and industrialization, global water consumption has increased by six times in the past century and is increasing at a rate of approximately 1% per year (WWAP[United Nations World Water Assessment Program]/UN-Water, 2018). However, social and economic activities discharge a large amount of domestic sewage and industrial sewage, causing a deterioration of water quality, destruction of the water environment, and fragile aquatic ecology. Meanwhile, extreme weather caused by climate change has led to frequent floods and droughts. Currently, the conflict between people and water is intensifying because of the complex and intersecting water issues of water shortage, environmental pollution, ecological degradation, frequent disasters, and lagging management (Xia et al., 2011). Thus, there is an urgent need to monitor water volume, water quality, and other water environment information in a macro, rapid, and economical way.

Indeed, remote sensing of water environment has become a good choice for this mission because it has an effective and continuous ability to monitor Earth's surface at multiple scales (Huang et al., 2018). At present, we have entered an unprecedented era of massive remote sensing data, and the available datasets are becoming diversified, mainly including massive real-time data acquired by satellites, drones, and other remote sensing equipments with different imaging methods and spatio-temporal resolutions (Amani et al., 2020). These datasets cover multiple observation sensors (e.g., optics, thermal infrared, microwave, LIDAR, fluorescence, and night lights) and have the typical “5V” characteristics of big data, namely large volume, variety, fast update velocity, veracity, and great potential value (David, 2008; Wouter & John, 2011). Quickly and effectively revealing the intricate connections between water and other earth variables, mining knowledge in various fields of water science, humanities, and social sciences, and realizing fine-grained, omni-directional, high-temporal, and multi-level simulations are the main advantages and values of water environment monitoring using remote sensing big data (Li et al., 2014).

Remote sensing big data for water environment monitoring is routinely described as a monitoring application for the water quantity and quality evaluated by water distribution and its quality characteristics, such as the concentration of chlorophyll-a (Chla), suspended particulate matter (SPM), and colored dissolved organic matter (CDOM). Remote sensing separates the atmospheric influence from the radiation received by the satellites/sensors and extracts the composition and characteristic information of water based on the remote sensing images. In recent years, the vigorous development of digital earth (Gore, 1998), big data, and cloud computing has promoted the extensive applications of remote sensing big data for water environment monitoring, such as water mapping (Alevizos, 2020; Chen et al., 2017; Ji et al., 2018; Murphy et al., 2018; Sagawa et al., 2019; Traganos et al., 2018; Wang et al., 2020; Zhou et al., 2019), quantitative estimation of key water quality parameters (Arabi et al., 2020; Cao et al., 2019; Chen et al., 2021; El Hourany et al., 2019; Eugenio et al., 2020; Griffin et al., 2018; Kõuts et al., 2007; Lin et al., 2018), flood and drought monitoring (Cretaux et al., 2011; Jiao et al., 2021), marine surveying and mapping (Chen et al., 2004; Choi et al., 2020; Feng et al., 2020; Ling et al., 2020; Prakash et al., 2021; Solanki et al., 2015; Wang et al., 2017; Xu et al., 2019; Zhang et al., 2014), and marine climate (Medina-Lopez & Urena-Fuentes, 2019).

Typically, these applications of remote sensing big data for water environment monitoring can be divided into two categories: water quantity and quality, with the key scientific issues of water extraction and quantitative estimation of water quality, respectively. To the best of our knowledge, however, apart from several reviews on remote sensing big data (Ma et al., 2015; Sudmanns et al., 2020), cloud computing platforms (Amani et al., 2020; Liang et al., 2018; Tamiminia et al., 2020), water extraction (Huang et al., 2018), drought monitoring (Jiao et al., 2021), and future hydrological directions (McCabe et al., 2017; Sun & Scanlon, 2019), there have been no systematic reviews of remote sensing big data for water environment monitoring from the two aspects of water extraction and the quantitative estimation of water quality.

To demonstrate the advancement of remote sensing big data for water environment monitoring, this paper reviews the latest progress and state-of-the-art methods in water extraction and the quantitative estimation of water quality using remote sensing. The current status, shortcomings and challenges, and recommendations and future directions of remote sensing big data for water environment monitoring are also detailed to advocate and encourage the ongoing exploration of the theory of remote sensing big data and its applications for water environment monitoring. The structure of this paper is as follows: (a) The image data source and model evaluation metrics are comprehensively presented in Section 2. (b) The state-of-the-art methods for water extraction are detailed in Section 3, including three categories of threshold-based methods, water indices, and machine learning-based methods. (c) The state-of-the-art estimation models for water quality are detailed in Section 4, including three categories of empirical models, semi-empirical/semi-analytical models, and machine learning-based models. On this basis, some shortcomings and three challenges of remote sensing big data for water environment monitoring are presented in Section 5. Correspondingly, in view of the current research deficiencies, scientific priorities, and difficulties, five suggestions are recommended in Section 6 to provide a potential roadmap and information support for researchers, practitioners, and management departments in the theoretical exploration and innovative application of remote sensing big data for water environment monitoring.

2 Remote Sensing Image Data Sources and Model Evaluation Metrics

In the past half century, the United States, Europe, China, Russia, Japan, Canada, India, Korea, and other nations or regions have operated numerous satellite systems that provide various remote sensing images and Earth observations to the scientific community, enterprises, and government departments (McCabe et al., 2017). According to the Union of Concerned Scientists Satellite Database (https://www.ucsusa.org/resources/satellite-database, Union of Concerned Scientists, 2021), there were 4,450 satellites orbiting the Earth in September 2021, with the United States, China, and Russia as the top three contributors, with 2,788, 431, and 167 satellites in orbit, respectively. Meanwhile, remote sensing big data for water environment monitoring based on a variety of satellite sensors has been developed rapidly. It is necessary to build a catalog of datasets and model evaluation metrics to facilitate a rapid response to the specific remote sensing applications. Therefore, the objectives of this section are as follows: (a) summarize the commonly used remote sensing images from three levels of coarse, medium, and high spatial resolution sensors, and their main parameters are presented in Table 1 (Baillarin et al., 2012; Coffer et al., 2020; Cook et al., 2001; Dial et al., 2003; Gomathi & Selvakumaran, 2018; Hu et al., 2012; Jiang et al., 2019; Kudela et al., 2019; Qi et al., 2017; Zhang et al., 2021); and (b) summarize the commonly used metrics for remote sensing methods from two aspects: water extraction and the quantitative estimation of water quality.

Table 1. Main Parameters of Typical Remote Sensing Images

Category	Sensor	Data availability	Height on orbit (km)	Orbital swath (km)	Spatial resolution (m)	Temporal resolution (day)	Bands	Spectral range(nm)	Signal-to-noise ratio	Acquisition method
Coarse resolution	AVHRR	1978∼	833–870	2,800	1,100	0.5	5	550–12,500	/	free
	MODIS	1999∼	705	2,330	250–1,000	0.5	36	400–14,400	/	free
	MERIS	2002–2012	790 ± 10	1,150	300	3	22	465–2,135	/	free
	GOCI	2010∼	35,837	2,500	500	1/24	8	402–885	545–945	free
	Sentinel-3	2016∼	814.5	1,270	300	2	21	400–1,020	50–168	free
Medium resolution	Landsat 1–3	1972–1983	907–915	185	78	18	4	500–1,100	<40	free
	Landsat-4/5	1982–2012	705	185	30–120	16	7	450–12,500	17–72	free
	Landsat-7	1999∼	705	185	15–60	16	8	450–12,500	13–78	free
	Landsat-8	2013∼	705	185	15–100	16	11	430–12,510	145–355	free
	Landsat-9	2021∼	705	185	15–100	16	11	435–12,500	162–442	free
	SPOT 1–4	1986∼2013	822	60	10–20	26	4–5	500–1,750	119–219	charge
	Hyperion	2000–2017	705	7.7	30	200	242	400–2,500	<50	free
	Sentinel-2	2015∼	786	290	10–60	5	13	420–2,300	50–174	free
High resolution	IKONOS	1999–2015	681	11.3	0.82–4	1.5–3	5	445–900	67–143	charge
	QuickBird	2001–2014	450–482	16.8–18	0.61–2.88	1–6	5	450–900	25–32	charge
	WorldView 1–4	2007∼	496	17.6	0.31–3.7	1.7–5.9	4–28	450–800	0.45–22	charge
	SPOT 5	2002–2015	822	60	2.5–20	26	5	480-1,750	/	charge
	SPOT 6/7	2012∼	694	60	1.5–6	26	5	500–890	/	charge
	ZY-3	2012∼	506	50	2.1–5.8	3–5	7	500–890	>25	charge
	GF-1/2/6	2013∼	631–645	45–90	0.8–16	1–5	5–13	450–900	34–294	free
	Zhuhai-1	2017∼	500	150	0.44–10	1–32	32	400–1,000	>300	free

2.1 Classification of Remote Sensing Image Data Sources

2.1.1 Coarse Spatial Resolution Remote Sensing Images

Although remote sensing sensors with coarse spatial resolution have inherent defects in the fine identification of small rivers, they have been widely used in water dynamic monitoring and water quality estimation because of their short revisit period, increased number of spectral bands, and higher signal to noise ratio (Hu et al., 2012; Huang et al., 2018; Kudela et al., 2019; Qi et al., 2017). The Advanced Very High Resolution Radiometer onboard the National Oceanic and Atmospheric Administration satellites (NOAA/AVHRR), Moderate Resolution Imaging Spectroradiometer (MODIS), Medium Resolution Imaging Spectrometer (MERIS), Geostationary Ocean Color Imager (GOCI), and Sentinel-3 OLCI are the most typical coarse resolution sensors.

These satellite sensors can be divided into global scale and regional scale. NOAA/AVHRR, MODIS and Sentinel-3 OLCI are excellent satellite sensors for remote sensing of water environment on a global scale. They have been widely used to estimate sea surface temperatures, Chla concentrations, and phytoplankton function types. The proportion and seasonal differences of phytoplankton with different functional types in the Mediterranean were estimated using integrated GlobColour data and AVHRR sea surface temperatures (El Hourany et al., 2019). Likewise, MODIS and Sentinel-3 OLCI have gained extensive attention worldwide because of their good imaging quality, high-frequency revisit, rich spectral bands, and free use. Considering these advantages, especially the rich hyperspectral information, various models have been developed to estimate water constituents such as Chla, SPM, and CDOM (Kravitz et al., 2020; Pahlevan et al., 2020). D. Liu et al. (2021) estimated the monthly water column-integrated algal biomass by comprehensively using the floating algae index, Chla concentration, wind speed, and MODIS Aqua images. The above three satellite sensors are excellent representatives of global-scale research. The GOCI satellite is an example of regional research on water environments and has been widely used in Japan, South Korea, and some parts of China where coverage is available (Choi et al., 2020; Feng et al., 2020; Park & Park, 2021; Tian et al., 2020). For example, Ling et al. (2020) developed a GOCI-based model to estimate the CDOM concentrations in the Bohai and Yellow seas.

2.1.2 Medium Spatial Resolution Remote Sensing Images

Medium-resolution remote sensing images (e.g., Landsat MSS/TM/ETM+/OLI/OLI2, SPOT 1–4, Hyperion, and Sentinel-2 MSI) are the most successful satellite sensors both in water extraction and the quantitative estimation of water quality because of their good harmony in temporal, spatial, and spectral resolutions. In particular, Landsat, as the first civilian Earth observation satellite, has been upgraded for nine generations since 1972. The stable image quality and suitable resolution between 15 m (panchromatic) and 78 m (near-infrared, red, green, and blue) make it ideal for various water dynamic monitoring, especially because all Landsat images have been freely available since 2008 (Huang et al., 2018). Pekel et al. (2016) used all the cloudless Landsat images between 1982 and 2015 to extract global water bodies and mapped and analyzed their interannual and monthly changes. Wang et al. (2017) developed a QRLTSS model based on Landsat images to retrieve the long-term changes of SPM concentration in the estuaries of China. Wang et al. (2004) retrieved the concentration of three water quality parameters (TOC, BOD, and COD) and estimated water quality changes in the reservoirs of Shenzhen, China, using Landsat TM images. Cao et al. (2020) applied an extreme gradient boosting tree to Landsat OLI and estimated the Chla concentrations of lakes in the lower reaches of the Yangtze River during 2013 and 2018, with higher accuracy (N = 102, R² = 0.79, and RMSE = 7.10 μg/L). Predictably, with the launch of the Landsat 9—the first mission of NASA/USGS Sustainable Land Imaging in September 2021—the Landsat images will be more widely used in water mapping and water quality estimation. Specially, Landsat-8 and Landsat-9 can work together to take a series of stable and predictable images across the entire earth every 8 days (Masek et al., 2020).

Compared with Landsat images, other medium-resolution images have their advantages and disadvantages. For example, SPOT 1–4 is a medium-resolution satellite with better spatial (10–20 m) resolution. However, the fee-based image acquisition policy restricts its widespread use in scientific research, enterprises, and management departments (Huang et al., 2018). In addition, Hyperion's 242 bands between 400 nm and 2,500 nm (spectral resolution up to 4 nm in some bands) promote its application in the quantitative estimation of water quality. Of particular importance is that Hyperion's high spectral resolution between 675 and 735 nm makes it very suitable for the quantitative estimation of Chla concentration. Chen et al. (2011) developed a three-band estimation model of Chla concentration suitable for turbid water bodies based on EO-1 Hyperion and synchronized in situ data, and successfully applied it to the Pearl River Estuary (R² = 0.64, RMSE = 2 mg/m³). Hicks et al. (2013) applied similar methods estimate a time series of turbidity and Chla concentration of shallow lakes of the Waikato Region, New Zealand. The capabilities of Hyperion have also been verified in mapping shallow bathymetry using multiple regression methods (Alevizos, 2020). However, the narrow width and low revisit period of Hyperion images limit their use in large-scale, high-frequency water environment monitoring. Sentinel-2, which was launched in 2015 by the European Space Agency, must also be mentioned. The 5-day revisit period, the spatial resolution of 10–60 m, and the advantages of free access, have enabled Sentinel-2 MSI to receive great attention since its release. Many researchers worldwide have used it to map the distribution of water bodies and estimate the critical water quality parameters. Chen et al. (2021) applied Sentinel-2 MSI images to estimate the Chla concentration in two large reservoirs and compared the advantages and disadvantages of the integrated machine learning algorithm (GA-ANN) and the traditional semi-analytical method. Likewise, based on Sentinel-2 MSI images, G. Liu et al. (2021) estimated the concentrations of CDOM and DOC in alpine lakes on the Qinghai-Tibet Plateau that are greatly affected by global climate change, and compared their spatio-temporal differences in freshwater lakes and saltwater lakes.

However, the above four types of medium-resolution optical images are all susceptible to clouds and rain, which limit their ability to capture instantaneous water environment scenarios and may miss key change nodes and driving factors, such as the monitoring of instantaneous flooding caused by typhoons. Furthermore, the longer re-entry period of the medium-resolution imaging (usually more than 10 days) may further exacerbate this problem.

2.1.3 High Spatial Resolution Remote Sensing Images

Water environment monitoring using high spatial resolution remote sensing images has been developing rapidly in the past 20 years because it has an irreplaceable role in identifying subtle changes of water system morphology, water depth, and water quality. IKONOS, QuickBird, WorldView, and ZY-3 with the meter, even sub-meter, level are the most commonly used images in water environment monitoring. For example, Ormeci et al. (2009) developed an accurately estimated model of Chla concentration using IKONOS(R² > 0.74, RMSE < 0.72 μg/L). Similarly, Liu et al. (2015) effectively estimated the total nitrogen concentration (TN) and total phosphorus concentration (TP) using both empirical model and artificial neural networks (ANN) model based on IKONOS with lower RMSE (0.89 mg/L) than using Landsat TM image (RMSE = 2.50 mg/L). Eugenio et al. (2020), Mortula et al. (2020), and X. Wang et al. (2018) also estimated the Chla concentration and further evaluated the eutrophication level of coastal lagoons, lakes, and reservoirs using WorldView images, respectively. In addition, other research demonstrated that IKONOS and Quickbird perform better than Landsat OLI in identifying different benthic coral reefs, which have a complex spatial heterogeneity and spectrum (Xu et al., 2019). However, the characteristics of narrow width, severe boundary distortion, and charges of these high-resolution satellite images partly limit their widespread use in scientific research and corporate sectors (Huang et al., 2018).

Notably, China has made significant progress in national and private high-resolution satellites in recent years and has taken the lead in realizing free, open high-resolution satellites such as GF-1/2/6 and Zhuhai-1. This has effectively promoted the application of high-resolution satellites in the high-precision monitoring of global water environments, especially in promoting the development of environmental satellites in countries along the Belt and Road. For example, Ma et al. (2021) developed a variety of high-precision retrieval models to effectively detect spatiotemporal changes of water transparency based on GF-1 images. Likewise, L. Liu et al. (2020) estimated the glacier mass loss and its contribution to river runoff in the source region of the Yangtze River from 2000 to 2018 using digital elevation models (DEMs) produced from ZY-3 tri-stereo scenes. Yin et al. (2021) developed an empirical model and quantitatively estimated the SPM concentration and water clarity in the Yuqiao reservoir using the Zhuhai-1 satellites. In addition, by comparing the classification results of wetland vegetation using GF-1, GF-2, ZY-3, Sentinel-2A, and Landsat-8 OLI, M. Liu et al. (2021) determined that high-resolution images can effectively reduce pixel mixing, and the classification accuracy increases with the improvement in the spatial resolution. In the future, with the continuous improvement in image resolution, refinement of the atmospheric window, and increase in spectral bands, mixed pixels, and the same spectrum of foreign objects may be effectively alleviated in remote sensing big data for water environment monitoring.

2.2 Model Evaluation Metrics

Suitable evaluation criteria are of great significance for model training, accuracy evaluation, over-fitting/under-fitting correction, and model transferability. While it is an exceptionally challenging task for most researchers to select representative evaluation indicators that can explain the research results from many evaluation criteria, there are several well-conducted and easy-to-understand metrics for image classification and quantitative remote sensing (Cheng & Han, 2016; Yaseen, 2021). Therefore, combined with the research progress of remote sensing big data for water environment monitoring, we review some model evaluation metrics, focusing on some indicators application to water extraction and the quantitative estimation of water quality.

For water extraction, the overall accuracy (OA), misclassification error (ME), omission error (OE), and ROC-AUC values are four universally approved metrics. The first three evaluation indicators can be represented in the confusion matrix. OA is equal to the sum of pixels correctly classified as water divided by the total number of water pixels, which is represented by the diagonal of the confusion matrix. ME refers to pixels that are classified as water, but actually belong to other surface types, which is denoted as the rows of the confusion matrix. OE refers to the number of pixels that belong to water but are classified as other surface types, which can be identified by the column of the confusion matrix. The ROC-AUC value (Hong et al., 2019) is the area under the curve of the ratio of the sensitivity (Equation 1) to the specificity (Equation 2). The value of ROC-AUC varies between 0.50 and 1; the larger the value, the better the performance, and when the value is greater than 0.70, the classification result is available.

(1)

(2)

where, TPR (true positive) and FNR (false negative) are the pixels that are correctly and incorrectly classified as water, respectively, and TNR (true negative) and FPR (false positive) are the pixels that are correctly and incorrectly classified as non-water bodies, respectively.

For the quantitative estimation of water quality, various metrics can be effectively used to evaluate the performance of the estimation model of water quality, including the coefficient of determination (R², Equation 3), mean absolute error (MAE, Equation 4), root mean square error (RMSE, Equation 5), mean square error (MSE, Equation 6), relative error (RE, Equation 7), residual prediction deviation (RPD, Equation 8), and confidence interval (CI, Equation 9). Suppose the observation data set has n observation points, is the average value, y_i and y_iʹ are the observed and predicted values of point i, respectively. The equations are as follows:

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Notably, although R² is the most commonly used and generally accepted evaluation indicator, excessive pursuit of a high value of R² can easily lead to over-fitting and insufficient portability of the model because R² is sensitive to extreme values but insensitive to the proportional difference (Tang et al., 2021; Yaseen, 2021). In practical applications, therefore, R² is often used in combination with RMSE, RPD, CI, and other indicators to balance the fitting accuracy, portability, and computational complexity. For example, Chen et al. (2021) and Song et al. (2013) demonstrated that R², RMSE, MAE, and RPD can be combined to develop a robust remote sensing estimation model of Chla concentration in various waters. According to Williams (2001), the model is very accurate when R² and RPD values are higher than 0.91 and 2.50. If RPD is higher than 2.00 and R² ranges from 0.82 to 0.90, the model can predict well. When RPD is higher than 1.50 and R² ranges from 0.66 to 0.81, the model can predict approximately. However, the prediction effect of a model is poor when R² is between 0.50 and 0.65 (Chen et al., 2021; Song et al., 2013). With the development of visual chart package libraries in programming languages, such as Python, R, and MATLAB, evaluation metrics aided by visual charts have been widely used, such as violin charts (Pena-Arancibia & Ahmad, 2021), heat maps (Huang et al., 2018; Song et al., 2013), box plots (Brisco et al., 2019; Tamiminia et al., 2020), and Taylor diagrams (Arabi et al., 2020; Zeng et al., 2015).

3 Classification of Methods for Water Extraction

The principle of water extraction is that the reflectance of water in the infrared (e.g., near-infrared band, NIR; short-wave infrared band, SWIR) channel is significantly lower than that of other land types (Figure 1). Many methods of water extraction have been developed based on this concept, providing geographic information support for global hydrology and water resources dynamic assessment and environmental evolution analysis. Feyisa et al. (2014), Li et al. (2021), and Huang et al. (2018) detailed the water extraction methods. Here, three categories of water extraction methods are summarized and compared their advantages and disadvantages: threshold-based methods, water indices, and machine learning-based methods.

Figure 1

Spectral reflectance curves of several land cover types, collected from the spectral library of ENVI software.

3.1 Threshold-Based Water Extraction

The threshold-based method, which generally refers to the single-band threshold method, is the simplest method for water extraction. Bartolucci et al. (1977) first proposed this method and applied it to the Landsat MSS. Subsequently, this method has been extensively studied for water extraction (Pekel et al., 2016; Wang et al., 2020). The primary theoretical basis of this method is that different ground objects have distinct reflection and absorption characteristics in the NIR band (Huang et al., 2018). For example, water bodies have strong absorption in the NIR band, whereas vegetation and soil show strong reflection characteristics. Therefore, it is possible to highlight the obvious difference between the water body and the background features (e.g., vegetation, minerals, and impervious surfaces) by setting an appropriate threshold to complete the separation of various objects.

As a single-band extraction method, the advantages and disadvantages of this method are evident. The algorithm is simple and easy to implement. However, it is difficult to apply the threshold value to other scenarios under different meteorological, spatial, and temporal conditions to achieve fine extraction of water bodies because the water's threshold value under different environments varies. For example, mountain water systems based on unified threshold extraction often have difficulty meeting the monitoring needs of water dynamic changes, especially for small rivers or lakes (Jia et al., 2018). The primary reason is that the radiance of the water body is various in different scenes because the difference between water and other land objects is easily affected by the shadow of the mountain and the surrounding environments like vegetation.

3.2 Index-Based Water Extraction

The water index is another simple and effective method to extract water, which is generally calculated from two or more bands to separate the water from various land covers. Since Crist (1985) proposed the Tasseled Cap Humidity (TCW) index, many water indices have been proposed and widely used in water extraction (Table 2). For example, McFeeters subsequently proposed the normalized difference water index (NDWI) to extract water using green and NIR bands in 1996. Likewise, the modified NDWI (MNDWI) was developed by replacing the bands of green and NIR in the NDWI with the bands of red and medium infrared (MIR), which successfully improved the defect that the NDWI cannot discern water bodies from construction land and can effectively identify wetlands and water bodies with different turbidity (Xu, 2006).

Table 2. Common Water Index Methods in Remote Sensing of Water Environment

Water index	Equation
TCW (Crist, 1985)
NDWI (McFeeters, 1996)
MNDWI (Xu, 2006)
OWLI (Guerschman et al., 2011)
AWEInon-shadow (Feyisa et al., 2014)
AWEIshadow (Feyisa et al., 2014)
WI2015 (Fisher et al., 2016)

Notably, these indices are not isolated, but can be nested with each other. Guerschman et al. (2011) proposed an open water likelihood index (OWLI) by integrating the NDWI, normalized difference vegetation index, SWIR band, and topographic relief index extracted from DEM data (MrVBF). And the effectiveness of extracting mountain water using OWLI has been significantly improved because this method reduces the uncertainty caused by terrain shadows. In addition, some new water indices suitable for specific satellites/sensors or study areas are still emerging rapidly. Feyisa et al. (2014) and Fisher et al. (2016) successively proposed water indices such as the automated water extraction index (AWEI_non-shadow) and its specific version with shadows (AWEI_shadow) and the 2015 water index (WI₂₀₁₅).

Overall, the water index method is the development of the single-band threshold method. Water indices use simple calculations to mine the spectral information of multiple bands, and then set the corresponding threshold to extract water body. When the used band is reduced to one, the water index method is single-band threshold method. Therefore, the advantages and disadvantages of water indices are similar to the single-band threshold method. In the future, it is necessary to develop additional novel water indices by combining machine learning methods to improve the extraction accuracy of the waters in shaded, snow, ice and seasonal water regions.

3.3 Machine Learning-Based Water Extraction

With significant advancements in semantic features and classifiers, many machine learning-based methods have been efficiently employed for water extraction and mapping (Amani et al., 2020; McCabe et al., 2017). Figure 2 presents a conceptual framework of machine learning-based water extraction, which can be performed by learning a classifier that collects the semantic features of water and classifies a set of calibrating images in a supervised or unsupervised framework (Cheng & Han, 2016). This framework is generally suitable for all possible land cover types, not just water. The input data are the images of the study area with the type label of land cover (e.g., water, vegetation, soil, and impervious surfaces), and the output data are their predicted labels, that is, water or non-water. In practice, semantic feature extraction (e.g., spectral characteristics, texture characteristics, and background environment), feature fusion, dimension reduction (e.g., Fourier transform, principal component analysis [PCA], redundancy analysis [RDA]), and classifier calibration jointly determine the accuracy and computational efficiency of these methods.

Figure 2

A conceptual overview of machine learning-based water extraction.

Random forest (RF), support vector machine (SVM, Qian et al., 2020), and ANN (Bui et al., 2021) are commonly used classifiers in water extraction. RF is a machine learning algorithm proposed by Ho (1995), and then formally proposed by Breiman (2001) based on decision tree (Y. Liu et al., 2021). RF performs decision tree modeling for each bootstrap resampled sample, and then obtains the final classification or prediction result through the combination of decision trees and voting, for example, water or non-water (Figure 3). Compared with the threshold-based method or water index-based methods, RF is more robust and accurate in water extraction because it has better tolerance for outliers (Labuzzetta et al., 2021). In the classification process of land cover, model parameters such as learning_rate, min_child_weight, gamma need to be adjusted to control the shrinking step size of the decision tree, the minimum leaf node weight, and the penalty coefficient of the loss function; meanwhile, it is generally necessary to set regularization parameters such as lambda and alpha to prevent overfitting. SVM is another intelligent method for classifying training samples by defining a hyperplane of the feature space class. The insensitive loss function, the slack variable and the penalty coefficient are introduced into the SVM to find a hyperplane that minimizes the "total deviation" of all sample points of the decision function, that is, the geometric distance from the sample point to the hyperplane is the smallest. Since it was first proposed by Vapnik and Vapnik in 1998, many researchers have promoted and improved the SVM algorithm owing to its advantages in fitting nonlinearity and pattern recognition. For example, Kadavi and Lee (2018) accurately distinguished four land cover types, including water, crop, vegetation, and snow, with a high classification accuracy of 98.08%. However, the classification accuracy of SVM is greatly affected by the loss kernel function, and the calculation is complicated and inefficient in some cases because it usually requires substantial cross-validation to find the best parameter settings. The ANN is a classifier of human-like thinking, which is composed of input data, neurons, hidden layers, activation functions, and output functions (Cheng & Han, 2016). In water extraction and other object identification, the ANN classifier performs well, and its improved versions with higher accuracy have rapidly emerged and are widely used. For example, Scarpa et al. (2018) performed the desired estimation of accurately extracting the features of various earth objects using the compact convolutional neural network (CNN).

Figure 3

Technical route of random forest.

Notably, single band threshold, water indices and machine learning-based methods do not exist isolation. When coupling three methods, the accuracy of water extraction can be improved significantly. For example, Tulbure et al. (2016) comprehensively used water indices and vegetation indices as semantic features to highlight water information, which was then input to RF model. The seasonal water bodies in the Murray Darling Basin were extracted with high accuracy, and the overall classification accuracy, production accuracy and user accuracy reached 99.9%, 87% and 96%, respectively. Likewise, Yang et al. (2015) integrated the fuzzy clustering method and a modified FCM to improve the accuracy of water extraction from heterogeneous backgrounds. The Kappa coefficient of classification is 0.94, which is significantly higher than the water indices (Kappa = 0.89) and SVM (Kappa = 0.89). Similarly, Labuzzetta et al. (2021) developed an integrated method SWIM that couples spatial clustering and RF, which improves the classification sensitivity and realizes the monthly extraction of surface water.

Although the automation, and classification accuracy of the above three categories methods can usually meet a majority of scenarios, several issues remain that need to be resolved urgently. First, the theory of mixed pixels needs to be strengthened, which is a prerequisite for the extraction of small water systems (McCabe et al., 2017). In response to this problem, many researchers have made useful explorations. For example, Jia et al. (2018) developed a new robust and low-cost spectral matching method based on discrete particle swarm optimization (SMDPSO), and verified it in many typical water bodies severely affected by shadows around the world with high Kappa coefficient and low standard deviation. The water extraction method for coupling optical images and SAR images is another major issue, which is the guarantee for seasonable, macroscopic, and uninterrupted monitoring of remote sensing (Cigna & Tapete, 2021). Because the current mature optical images are easily covered by cloudy and rainy, the available remote sensing images are insufficient. Many researchers have made useful discoveries. For example, Saghafi et al. (2021) developed a coupling machine learning-based method to obtain higher spatial accuracy and time frequency water mapping by combining multiple machine learning methods and optical (Landsat OLI and Sentinel-2 MSI) and Sentinel-1 SAR sensors. As mentioned in Section 2.2, in addition, choosing the appropriate model evaluation criteria that helps to evaluate the model more objectively and quantitatively is another problem that needs to be solved, thereby providing a reference for the calibration and optimization of the model (Yaseen, 2021).

4 Remote Sensing Estimation Methods for Water Quality

The radiation information L_sw received by the remote sensing sensors (Figure 4) mainly includes the water-leaving radiance L_w from the reflection and scattering of the water and its bottom material, the radiance r*L_sky of the atmosphere and water vapor, the radiance L_g of specular reflections on the water surface, solar flares, etc. The relationship between L_g, L_w, r*L_sky, and L_sw is shown in Equation 10. The theoretical basis of water quality estimation is to construct the imaging relationship between the true out-of-water reflectance R_rs(λ) of each water component, the water-leaving radiance L_w and the spectral channels of the remote sensing sensors. Furthermore, the inherent optical components of water are estimated by remote sensing images. Among them, the accurate out-of-water reflectance R_rs(λ) is the key and prerequisite for the quantitative estimation of water quality parameters. Assuming that the upward irradiance and downward irradiance on the surface of the water are L_u(λ,0⁺) and E_d(λ,0⁺) respectively, then the relationship between R_rs(λ) and L_w, r*L_sky, L_u(λ,0⁺) and E_d(λ,0⁺) is shown in Equation 11 (Lee & Carder, 2004). Notably, since the water-leaving radiance L_w of water cannot be directly measured, and the sky light r*L_sky has a great influence on the value of R_rs(λ), it is necessary to remove noises (e.g., the atmosphere, solar flares) from the upstream radiation information L_u received by the sensor before the quantitative estimation of water quality (Dall'Olmo & Gitelson, 2006).

Figure 4

Schematic diagram of the interaction between the main optically active substances in the atmosphere and water and the remote sensing sensors (Dornhofer & Oppelt, 2016).

Although water components and their optical properties are complex, the optically active substances that determine the reflectance from water mainly include pure water, phytoplankton, non-pigmented suspended matter, and colored soluble organic matter (Figure 4, Dornhofer & Oppelt, 2016). These optically active substances affect water's R_rs(λ) through absorption and scattering. Among them, the optical information that does not change with the incident light source and angle is Inherent Optical Properties (IOPs); conversely, the optical information that changes with the incident light source, incident angle, and diffuse attenuation coefficient is Apparent Optical Properties (AOPs). The goal of the quantitative estimation of water quality is to establish the relationship between the radiance received by remote sensing sensors and water's R_rs(λ) to estimate water constitutes using their optical properties.

(10)

(11)

In the past half century, numerous methods have been developed to estimate the water quality from various satellite images. These methods can be broadly divided into two categories: physics-driven methods and data-driven methods. Physics-driven methods can be further refined into radiation transfer models and data assimilation; accordingly, data-driven methods can also be refined into empirical models, semi-empirical/semi-analytical models, and machine learning-based models. Notably, these methods are not applied in isolation in the process of environmental monitoring. In fact, remote sensing monitoring is carried out for specific applications and problems in most cases, which requires the integration of the above-mentioned multiple methods for cross-use.

4.1 Remote Sensing Estimation Models Based on Physics-Driven Methods

Physics-driven methods are mainly to simulate the radiation transmission process of water and gas in different optical scenes according to the composition of the water, and then calculate its imaging characteristics on the remote sensing images. Radiation transfer models and data assimilation are two major categories of physics-driven methods. Due to the consideration of physical conservation, momentum conservation, and inherent absorption and scattering characteristics of water components, physics-driven methods have clear physical meaning, rigorous causality, and strong portability, and have been widely used in the quantitative estimation of water quality.

4.1.1 Radiative Transfer Model

The radiative transfer model is a popular approach for water quality remote estimations. Several different numerical methods of the radiative transfer equation have been developed such as the matrix-operator method, the discrete ordinates method, the spherical harmonic method, the Monte-Carlo solutions method (He et al., 2010). Among them, the radiation transfer equation between AOPs and IOPs developed by Maul (1985) and Mobley (1994) has been widely used.

Currently, there has been a considerable amount of research on this method. For example, Arabi et al. (2020) retrieved the daily changes of various water components (Chla, SPM, and CDOM) from 2013 to 2018 using radiation transfer equation and multi-source satellite images including MERIS, Sentinel-2 MSI, and Sentinel-3 OLCI, providing scientific and technological support for water quality management and abnormal monitoring of the Wadden Sea. He et al. (2010) developed a vector radiative transfer model termed PCOART for the coupled ocean–atmosphere system. However, the defects of radiation transfer equation are also obvious, such as many variables, complex computation, and difficulty in measuring the absorption and scattering coefficients of water. Indeed, many studies have made many simplifications and improvements to the radiation transfer equation in applications based on certain assumptions. For example, Eugenio et al. (2020) tuned and reconciled the multichannel radiative transfer model and atmospheric correction algorithms to estimate Chla concentrations using multispectral (Worldview satellite) and hyperspectral (drone and airborne) images in a natural reserve of Maspalomas, Spain. Another simplified method is the semi-analytical method, also called the semi-empirical method, which will be detailed in Section 4.2.2.

4.1.2 Estimation Models Combined With Data Assimilation

Estimation models combined with data assimilation are another type of popular physics-driven methods for the quantitative estimation of water quality (Jensen et al., 1989; Kõuts et al., 2007; Mano et al., 2015; Pleskachevsky et al., 2005). Benefiting from numerous available remote sensing images, many numerical models originally designed for specific applications have been subsequently utilized to produce new capabilities (Jiao et al., 2021). For example, Jensen et al. (1989) mapped the SPM concentration distribution to help understand and manage complex physical processes in coastal lagoons by integrating numerical simulation and remote sensing retrieval results. Likewise, Kõuts et al. (2007) evaluated the water quality of Pakri Bay, the southern Gulf of Finland, using satellite RS and three numerical simulation models, including the hydrodynamic model, particle transport model, and benthic macro algae growth model. Together with physical and causal models, exploring novel applications of data assimilation provides more opportunities for water environment monitoring.

Figure 5 illustrates a conceptual diagram of water quality estimation using data assimilation. The three key scientific issues involved in this method are as follows: (a) the development, benchmarking, and optimization of remote sensing retrieval models for key water ecology parameters, (b) data fusion and scale consistency check of multi-sensor remote sensing images, and (c) pattern matching and data assimilation of different numerical simulation models. In practice, these methods use high-precision remote sensing retrieval results as a data-driven field of numerical simulation model such as ECOMSED and Delft 3D (Chen et al., 2010; Zhang et al., 2015). Through the calibration of turbulence, convection, and other parameters, it activates the long-term and large-scale physical simulation of water quality on a three-dimensional scale. In addition, other physics and hydrodynamic models also provide opportunities to supplement remote sensing data that can respond to discrete states at different points in time (Chen et al., 2004; Jensen et al., 1989; Kõuts et al., 2007; Pleskachevsky et al., 2005). The FVCOM and 3D-COHERENS, which were originally designed for coastal sea areas, have illustrated the potential for environmental monitoring of water quality in inland waters (Mano et al., 2015).

Figure 5

Conceptual diagram of water quality estimation using data assimilation.

Compared with radiation transfer equation, the integrated method combined remote sensing retrieval of water quality with data assimilation has significant advantages in mapping the 3D distribution of water quality and overcoming the unavailability of images when covered by clouds and rain, providing theoretical support for long-term, multi-scale water quality monitoring. For example, Chen et al. (2010) coupled MERIS images and ECOMSED (a hydrodynamic, wave, and sediment transport model) to simulate the transport process of SPM in the Bohai Sea. Likewise, Delft3D-FLOW is also integrated with various remote sensing images, such as MODIS, and has been widely applied in various waters, including oceans, reservoirs, and lakes (Zhang et al., 2014, 2015). Solanki et al. (2015) developed a coupling method for data analysis and bio-physical processes, which effectively mapped the surface profiles to understand the linkage of the sea surface temperature, Chla concentrations, and surface height anomalies with marine fishery resources. In addition, based on the theoretical knowledge of fluid mechanics, conservation of momentum, and conservation of mass, these models use the measured point data for numerical simulation to compensate the lack of images caused by rainy weather. Thus, leveraging various numerical models of both physics and causality with multi-source remote sensing offers complementary data and acquires insights to better understand the evolution characteristics of various water environments.

Notably, this method has shortcomings. First, the calculation is complicated, and model runs take a long time. In the actual operation and calibration process, it is often necessary to interpret the complicated partial differential equations. Second, the final simulation effect of the model is greatly affected by the estimation results of the quantitative remote sensing, which requires a high accuracy of the remote sensing estimation. Furthermore, this method requires a substantial amount of hydrodynamic data and high-precision bottom terrain data, and it has decreased effectiveness on river sections with insufficient data or dam interference. In addition, the spatial scale effect between the estimation result of the remote sensing and the numerical simulation is significant. When the matching degree between the two methods is not good, the simulation result may be inaccurate, and the results are difficult to interpret.

4.2 Remote Sensing Estimation Models Based on Data-Driven Methods

With the rapid increase in available remote sensing images and the extensive use of advanced information technologies, such as machine learning, cloud platforms, and data mining in remote sensing of water environment, data-driven methods have been widely used in remote sensing big data for water environment monitoring owing to the low input parameters and simple calculation, but the simulation results have relatively high accuracy. This section provides a detailed overview of data-driven methods from empirical methods, semi-empirical/analytical methods, and machine learning-based methods.

4.2.1 Empirical Models

Empirical models have been widely used in the quantitative estimation of water quality, such as three water color parameters (Chla, SPM, and CDOM), water temperature, water depth, and biochemical parameters (Griffin et al., 2018; Li et al., 2021; Mortula et al., 2020; Ormeci et al., 2009; Prakash et al., 2021; Song et al., 2017; Sun & Scanlon, 2019; Traganos et al., 2018; Wang et al., 2004; Wu et al., 2013; Yin et al., 2021). Table 3 reports some quantitative estimation studies of water quality using empirical models. Most of these methods are simple linear, exponential, or logarithmic models, which are often based on the statistical relationship between the target parameters of the water body and the reflectance of remote sensing images. For example, Griffin et al. (2018) developed a higher linear model to remotely estimate the CDOM concentration in the Arctic Ocean drainage basin using long-term Landsat images from the Google Earth Engine (GEE). Similarly, D. Liu et al. (2021) developed a process-oriented method to estimate the water column-integrated algal biomass using MODIS/Aqua images. Undoubtedly, the advantages and disadvantages of this method are prominent in practice. The advantages are easy to understand and are maneuverable. However, these methods lack the theoretical basis, unstable statistical relationships, poor repeatability, and difficult popularization. Indeed, although the empirical models had higher R² and lower RMSE values for the calibration model of their study areas, these models are difficult to achieve stable high accuracy in other study areas.

Table 3. Previous Empirical Models of Estimating Water Quality Using Remote Sensing

References	Data	Developed models	Performance metrics	Atmospheric correction methods	Study area	Research remark
Griffin et al., 2018	Landsat TM/ETM+/OLI	LOADEST, linear model using three bands	R^2, RMSE, MAPD	GEE, 6S	The Arctic Ocean drainage basin	Developed a Landsat-based approximations of CDOM using GEE
D. Liu et al., 2021	MODIS/Aqua	Process-oriented algorithms	R^2, r, P-value, bias, MAPD	Rayleigh-corrected reflectance	Lakes Chaohu and Taihu, China	Developed a process-oriented method to estimate the water column-integrated algal biomass using MODIS/Aqua images
Mortula et al., 2020	WorldView-2, Landsat ETM+	Band ratios	R^2, RMSE, P-value, CI, Residual	The Cosine of the Sun Zenith Angle (COST) algorithm	Dubai creek	Developed a remote estimation model of Chla concentration, TN/TP ratio to assess the eutrophication level
Ormeci et al., 2009	IKONOS	Multivariate regression model	R^2, RMSE	ATCOR 2	Istanbul, north-west Marmara, Turkey.	Developed an effectively and easily model to map Chla concentration.
Prakash et al., 2021	Landsat OLI	Linear model	Percentage reduction, RMSE	The Dark Spectrum Fitting algorithm in ACOLITE	Chennai, India	Analyzed the SPM concentration and the impact of the COVID lockdown on coastal water across the Chennai region
Song et al., 2017	Landsat TM/ETM+/OLI, MOD09GA	Linear model of two band ratios	R^2, RMSE, MRE	Dark-object subtraction (DOS) method	Lakes and reservoirs in Northeast China	Developed the PAR estimation model which is suitable for Landsat and MODIS
Traganos et al., 2018	Sentinel-2 MSI	Empirical Satellite-Derived Bathymetries	R^2, RMSE	GEE	Four waters in Greece including National Park and the Aegean Sea	Proposed a complete preprocessing chain of remotely mapping the bathymetry using GEE
Wang et al., 2004	Landsat TM	TOC = 6.41– 85.29 × ρ1 - 2.05 × ρ2 - 24.96 × ρ3; BOD = 1.79 – 0.789 × ρ1 + 52.36 × ρ2 - 3.28 × ρ3; COD = 2.76 –17.27 × ρ1 + 72.15 × ρ2 -12.11 × ρ3	r, P-value, MRE	Improved dark-object subtraction technique for atmospheric haze correction	Reservoirs in Shenzhen, China	Developed the retrieval models of TOC, BOD, and COD based on Landsat TM images
Wu et al., 2013	Landsat TM, MODIS	Exponential, polynomial and linear models	R^2, RMSE, RRMSE, r, P-value, intercept, slope	/	Poyang lake, China	Proposed an integrated approach for estimating the concentration of SPM
Yin et al., 2021	Zhuhai-1	Linear model of two band ratios	R^2, RMSE	/	Yuqiao reservoir, China	Explored the preliminary study of Zhuhai-1 hyperspectral images on estimating water quality

4.2.2 Semi-Empirical/Semi-Analytical Models

The semi-empirical model (also called semi-analytical), a compromise between the radiative transfer model and the empirical model, is another popular data-driven methods for quantitative estimation of water quality. Table 4 presents previous research on semi-empirical models of the quantitative estimation of water quality. The study area of these studies involves the open ocean (Ling et al., 2020), estuaries (Chen et al., 2011; C. Wang et al., 2018), lakes (Ma et al., 2021; Song et al., 2020), reservoirs (Zhao et al., 2020), and rivers (Kravitz et al., 2020). In addition, these studies usually focus on the quantitative estimation of optical characteristic parameters of water quality, such as Chla concentration (Kravitz et al., 2020; G. Liu et al., 2020), SPM concentration and transparency (Ma et al., 2021; Song et al., 2020; Wang et al., 2017; Zhao et al., 2020), and CDOM (Ling et al., 2020; G. Liu et al., 2021). For example, the three-band model (Equation 13) for Chla developed by Dall'Olmo and Gitelson (2006) is a typical semi-empirical model, which has been widely used owing to its high accuracy and robustness (Chen et al., 2011; Song et al., 2012). The three-band model is based on the assumption that the water reflectance in the near-infrared band is zero, however, the estimation results are partially affected by the back reflection and absorption of suspended objects in the near-infrared band. Subsequently, the researchers introduced the fourth characteristic band based on the three-band model, which effectively suppressed the strong absorption of pure water near 735 nm and the backscattering of suspended matter in the near-infrared band (Le et al., 2009, 2011). Due to the strong absorption near 675 nm and the high backscattering and low absorption near 700 nm of Chla, the band ratio and spectral derivative near the red edge of 675∼710 nm have been widely used to estimate the concentration of Chla (Chen et al., 2021; Gitelson et al., 2008; Neil et al., 2019).

Table 4. Previous Semi-Empirical Models of Estimating Water Quality Using Remote Sensing

References	Data	Developed models	Performance metrics	Atmospheric correction methods	Study area	Research remark
Chen et al., 2011	ASD spectra, Hyperion	331.01 × [R^−1(684)-R^−1(690)] × R(718)+14.609	R^2, RMSE	FLAASH	The Pearl River Estuary, China	Evaluated the applicability of the three-band model and Hyperion sensor for Chla concentration estimation
Kravitz et al., 2020	Sentinel-3 OLCI, MERIS	Derivative models, band ratios, band difference models	R^2, RMSE, Bias, RE	MODTRAN, 6SV1	Four dams of Hartebeespoort, Roodeplaat, Bronkhorstspruit and Vaal in South Africa	Compared the pros and cons of various sensors, atmospheric correction methods and retrieval models
Ling et al., 2020	Hyper-Profiler spectra, GOCI	Single bands, band ratios, and other band combinations	R^2, RMSE, MAE, MRE, Bias, RE	GOCI Data Processing System	Bohai Sea and Yellow Sea	Developed and validated remote sensing methods to estimate the CDOM concentration using GOCI images
G. Liu et al., 2020	Sentinel-3 OLCI, MERIS, VIIRS/NPP, Field spectra	Improved Quasi-Analytical Algorithm (QAA)	R^2, MAPD, MND, NN, RMSE	SNAP 7.0, SeaDAS 7.5	36 waters including coastal waters, inland lakes, reservoirs and rivers	Developed a Chla concentration estimation model suitable for various turbid waters
G. Liu et al., 2021	ASD spectra, Sentinel-2 MSI	Linear model of two band ratios	R^2, P-value	Sen2Cor	Honghe National Nature Reserve, China	Developed the remote estimation models of CDOM and DOC using Sentinel-2 satellite
Ma et al., 2021	ASD spectra, GF-1	Linear regression	R^2, SD	Unknown	Shahu Lake, Third Drainage, Bird Island and Old Wharf	Compared pros and cons of semi-empirical models and empirical models
Song et al., 2020	ASD spectra, Landsat TM/ETM+/OLI	Linear regression	R^2, RMSE, MAE, Slope, Intercept	TOA, GEE, Rayleigh-corrected, Dark Spectrum Fitting	2759 lakes with area >1 km^2 distributed in five limnetic regions of China	Developed remote estimation model of lake clarity which suitable for long-term changes monitoring of Secchi disk depth at national or continental scale
Wang et al., 2017; C. Wang et al., 2018	ASD spectra, Landsat TM/ETM+/OLI, Hyperion	QRLTSS model	R^2, RMSE, MRE	6S	Estuaries in China	A wide range TSS (4.3–577.2 mg/L) model was developed
Zhao et al., 2020	ASD spectra, Landsat OLI	Three-band model	R^2, RMSE, MRE	FLAASH	Hedi reservoir, China	Developed an integrated model combining Markov and remote sensing to estimate the SPM concentration

Overall, this method generally has both physical and statistical significance, and the prediction results are relatively credible. In practice, assuming that some IOPs are fixed, researchers have further developed an estimation model of water quality parameters by measuring and analyzing the imaging mechanism of some IOPs and their water radiance, such as attenuation coefficient of water, and backscattering coefficient. Notably, although the semi-empirical model improves the physical meaning and universality compared with empirical methods, the spatio-temporal limitations of this method are still significant. In the future, it will be necessary to increase research on specific areas and specific water constituents to promote the innovative application of remote sensing big data for water environment monitoring.

The mathematical forms of empirical and simi-empirical models are diverse, mainly including logarithmic form (Equation 12), exponential form (Equation 13), Gordon form (Equation 14, Gordon & Wang, 1994), band ratio (Equation 15), Normalized Difference Water Quality Index (Equation16), three-band model (Equation 17, Chen et al., 2021; Gitelson et al., 2008) and four-band (Equation 18), etc.

(12)

(13)

(14)

(15)

(16)

(17)

(18)

where, WQ is the water quality parameter, and R(λ_i) is the reflectance in spectral band λ_i. Theoretically, R(λ₁) and R(λ₂) are the wavelength bands where the sensitivity to Chla and other water components are the greatest and the least, while the absorption of other components at λ₁ and λ₂ is equal. The difference between R(λ₁) and R(λ₂) is related to the absorption at λ₁. Therefore, R(λ₃) or [R(λ₄)⁻¹-R(λ₃)⁻¹] is introduced to further eliminate the back reflection and absorption of suspended matter in the near infrared. In practice, an iterative method was used to select the optimal band sets for λ₁, λ₂, λ₃ and λ₄. In the three-band modeling of Chla, we fixed the position of any two-band parameters, and regressed equation (Equation 17) to approximate the position of the third band, according to the maximal correlative coefficient of Chla estimation or minimal RMSE of Chla estimation. The above steps were repeated until each optimal band was constant, and the correlative coefficient of [R (λ₁) ⁻¹-R (λ₂) ⁻¹] × R(λ₃) versus Chla was also unvarying. The last band positions were model-calibrating wavelengths for Chla estimation using Least Squares Linear Regression method (Chen et al., 2011, 2021). With the rapid development of programming languages such as Python, R and MATLAB, as well as the improvement of computer performance, the development and calibration time of estimation models has been greatly shortened, providing technical support for the development of more complex forms of semi-empirical models.

4.2.3 Machine Learning-Based Models

With the rapid increase in available remote sensing images and real-time Earth observation demand, the direction of remote sensing big data for water environment monitoring has gradually changed (Cretaux et al., 2011; Gore, 1998), including changes such as: (a) Focus on dense-time water environment evolution instead of random or sparse sampling; (b) give more insight into the water environment evolution trend on a macro scale; (c) focus on multi-factor correlation analysis to solve scientific problems. To meet these changes, machine learning-based models for water environment monitoring has recently received remarkable attention (McCabe et al., 2017). Table 5 presents some machine learning-based models for estimating the water quality using remote sensing.

Table 5. Previous Machine Learning-Based Models of Estimating Water Quality Using Remote Sensing

References	Data	Developed models	Performance metrics	Atmospheric correction methods	Study area	Research remark
Alevizos, 2020	Hyperion	RF and kNN	R^2, RMSE	FLAASH	Caribbean sea	Developed machine learning-based methods and strategies for evaluating bathymetry
Cao et al., 2020	Landsat OLI	Extreme gradient boosting tree (BST, XGBoost)	R^2, RMSE, MAE, bias, MAPE, slope	Pseudo reflectance products	Inland lakes in the lower reaches of the Yangtze River in China	Developed the BST algorithms to estimate the concentration of Chla using Landsat images
Chen et al., 2004	Landsat TM, SeaWiFS, NOAA/AVHRR	Maximum likelihood, neural network, and SVM	ANOVA	COST method	The Pearl River Estuary, China	Compared three machine learning methods to rapidly detect the changes of coastal water quality
Chen et al., 2021	ASD spectra, Sentinel-2 MSI	GA-ANN	R^2, RMSE, MAE, RPD, Slope, Intercept	Sen2Cor	Coastal area of Western Guangdong, South China	Evaluated and expanded the application of GA-ANN in Chla estimation
Feng et al., 2020	GOCI	NN	R^2, RMSE, MRE	GOCI Data Processing System	The Yellow River Estuary, China	Developed an NN method which is suitable to estimate the water turbidity of coastal area
Liu et al., 2015	IKONOS	NN	R^2, RMSE	FLAASH	Small urban rivers and lakes in China	Compared the suitability of multiple linear regressions and ANN to estimate the concentration of TN and TP
Medina-Lopez & Urena-Fuentes, 2019	Sentinel-2 MSI	DNN	r, MAE	Raw satellite data	Global ocean	Develop the retrieval models of sea surface salinity and temperature using raw satellite data
Pahlevan et al., 2020	Sentinel-2 MSI, Sentinel-3 OLCI	Mixture density network	R^2, RMSE, MAE, Bias	SeaDAS, POLYMER, ACOLITE	Inland and coastal waters	Explored a preliminary study on seamless construction of Chla concentration in inland and coastal waters
Park & Park, 2021	GOCI	DNN	ROC-AUC	GOCI Data Processing System	Yellow Sea and East Sea	The MFNN method was applied to eliminate erroneous effects on estimating the Chla concentration
Sagawa et al., 2019	Landsat OLI	RF	R^2, RMSE, ME, CI	GEE	Hateruma and Taketomi, Japan; Oahu, USA; Guanica, Puerto Rico; Efate, Vanuatu	Developed a RF-based estimation method of shallow water bathymetry
Song et al., 2014	AISA hyperspectral image	GA-PLS	R^2, RMSE, rRMSE, MAE, RPD	/	Morse Reservoir, IN, USA	The eutrophication level was evaluated through the quantitative retrieve of Chla, PC, TSM, and SDD
X. Wang et al., 2018	WorldView-2, MERIS	Two-band model, three-band model, stepwise regression algorithm (SRA) and SVM.	R^2, RMSE, URMSE	FLAASH, MODTRAN	Guanting Reservoir, China	Verified and compared the cons and pros of several Chla retrieval models
Zhu et al., 2020	Landsat OLI	Sigmoid model, two band combinations	R^2, RMSE, RRMSE, AME	FLAASH, 6S, and ACOLITE	Zhoushan Islands, China	Verified and compared the cons and pros of several SPM retrieval models

Essentially, this method simulates the thinking process of the human brain and transforms the water quality estimation problem into a computer-recognizable input method. The advantages of machine learning in nonlinear regression fitting have promoted its widespread use in quantitative remote sensing. Many machine learning algorithms, such as partial least squares regression (PLSR, Song et al., 2014), support vector regression (SVR, X. Wang et al., 2018), random forest regression (RFR, Alevizos, 2020; Du et al., 2018; Nguyen et al., 2018; Sagawa et al., 2019), XGBoost (Cao et al., 2020; Cui et al., 2021; Nguyen et al., 2021; Tiyasha et al., 2021), ANN (Chen et al., 2004; Chen et al., 2021; Feng et al., 2020; Liu et al., 2015; Medina-Lopez & Urena-Fuentes, 2019), and long short-term memory (Mohebzadeh & Lee, 2021; Sagan et al., 2020; Yu et al., 2020), have been widely used to estimate water environment characteristics, including Chla, SPM, sea surface salinity, and temperature. While ensuring that the spectral characteristics are highly correlated with water quality parameters, PLSR maximizes the variance of the main components of sensitive bands and water quality parameters. PLSR has the advantages of fewer setting parameters, easy identification of data noise, and easy interpretation of principal component analysis regression coefficients. However, PLSR is more sensitive to outliers; and when the training data does not have an inverse matrix, PLSR needs to be optimized in combination with other methods. SVR introduces Lagrangian multiplier and kernel function to map the nonlinear relationship space between spectra and water quality parameters to high-dimensional feature space. Subsequently, according to the Kuhn-Tucker theorem, the regression fitting of water quality parameters is realized by solving the calculation process of partial differential and duality. Similar to water extraction, SVR has the significant advantage in the fitting of small samples; but the simulation accuracy of SVR is greatly affected by the loss kernel function. RFR and XGBoost are two methods based on integrated decision trees that map and minimize the information entropy between input variables and target variables. RFR and XGBoost algorithms have similar technical route and advantages including easy to implement, low computational overhead, and powerful performance in many simulating tasks. The main difference is that XGBoost uses the second-order Taylor formula to expand and optimize the loss function of the decision tree and the objective function of the previous tree. XGBoost has other advantages, such as powerful regularization promotion, advanced pruning strategy, good parallel processing and scalability, high degree of flexibility and built-in default value cleaning. However, the integrated decision tree may overfit due to noise-induced pruning failure. ANN has the function of realizing complex nonlinear mapping, which is especially suitable for solving problems with complex internal mechanisms. However, the objective function to be optimized by ANN is more complicated, and it has the shortcomings that the gradient descent method may produce local optimal solutions and over-fitting.

In practice, water composition estimation can be performed by training a model that captures the variations in composition appearances and imaging characteristics from a set of calibration data in a supervised or unsupervised framework (Cheng & Han, 2016). The input data of the learner is a set of water images represented by imaging features such as spectra and texture, and the output is the corresponding prediction element, that is, the concentrations of Chla, SPM, and CDOM, etc. In this process, the learning operator, activation function, feedback function, and convergence criterion jointly determine the accuracy and operating efficiency of the algorithm. For example, in ANN modeling, hidden layers, activation function, and the number of neurons are three key parameters controlling its performance and complexity (Alevizos, 2020; Medina-Lopez & Urena-Fuentes, 2019).

Comparing the three commonly used activation functions of ANN, including Sigmoid, ReLU, and Softmax, in estimating the concentration of Chla, the predicting accuracy and model stability exhibit the following law: ReLU > Sigmoid > Softmax (Chen et al., 2021), and the possible reasons are as follows. (a) For the deep network, the gradient disappears easily when the sigmoid function is propagated back. For example, when it is close to the saturation zone, the transformation is too slow and the derivative tends to zero, which causes a loss of information and disappearance of the gradient (Oostwal et al., 2021). Moreover, the output of the sigmoid function is not zero-centered (Kocak & Siray, 2021). When the data of the input neuron is positive, the gradient of the sigmoid function will be all positive or all negative in the process of backpropagation, which will cause a z-shaped drop when the gradient descent weight is updated. Therefore, a deep network calibration using a sigmoid function cannot be completed effectively. (b) In contrast, ReLU has unilateral inhibition and a relatively wide excitability boundary. The gradient of ReLU is 1, which is close to the working principle of the human neurocortex, and there will be no problems of gradient disappearance or gradient explosion such as sigmoid function. (c) The softmax function fails to obtain a good estimation of Chla concentration because the activation function of softmax is primarily designed for the logistic regression of more than one output neuron (Gong et al., 2020). Softmax is used for multi-class neural network output by ensuring that the sum of all output neurons is 1.0. It is not surprising that the ANN model using the softmax function cannot fit the complex non-linear relationship between the Chla concentration and spectral information.

Another advantage of machine learning-based models is that they can simulate the complex water environment phenomena in the natural world to realize the high-precision estimation of water components, while the calibrated model can quickly estimate multiple water elements simultaneously. For example, Medina-Lopez and Urena-Fuentes (2019) achieved the simultaneous estimation of multiple water quality parameters, providing data support for the accurate assessment of water eutrophication status. However, as a typical data-driven method, the accuracy and portability of machine learning-based models are significantly affected by the calibrating samples. Therefore, the calibrating samples for the model should not only reduce the errors but also cover the possible numerical ranges as much as possible.

5 Shortcomings and Challenges

5.1 The New Data Gap Caused by Massive Heterogeneous Data

With the continuous emergence of higher (high resolution), faster (short revisit period), and finer (more detailed atmospheric window and higher spectral resolution) remote sensing data, remote sensing big data provides the opportunity to quickly and effectively reveal the intricate connections between water environment elements, mine varied hydrological, environmental, and ecological knowledge, and achieve fine-grained, omni-directional, high-temporal, and multi-level simulations of water ecosystems. The current processing methods and models have difficulty in processing the massive remote sensing big data with heterogeneous formats (e.g., sensors, resolution, and storage format), however, resulting in a new generation of data gap. Common file formats of remote sensing images include img, dat, Tiff, ASCII, Grid, NetCDF, and HDF (Sun & Scanlon, 2019). Overall, scientific research on remote sensing big data is still in its infancy, and the engineering technology research of big data is at the forefront of this scientific research (Li et al., 2014).

How to integrate, process, and analyze remote sensing big data in a timely and effective manner, and then mining useful information to solve the water environment problems in the process of global warming and urbanization will be the focus and difficulty in the future. Fortunately, with the continuous increase in the application of water environments in various specific directions (such as data fusion, spatial downscaling, data mining based on cloud computing, etc.), remote sensing of water environment has become the focus of remote sensing big data and hydrology, providing critical data and solutions for water resource management, such as water extraction, water pollution monitoring, water ecological assessment, flood disaster assessment, and global climate change (Cao et al., 2019; Chen et al., 2017; Griffin et al., 2018; Ji et al., 2018; Lin et al., 2018; Murphy et al., 2018; Pekel et al., 2016; Sagawa et al., 2019; Traganos et al., 2018; Wang et al., 2020; Zhou et al., 2019). In the future, the storage and calculation of massive heterogeneous remote sensing data and specific research on water environment monitoring should be further strengthened.

5.2 Inefficient Water Environment Monitoring Due To “Low Spatiotemporal Resolution”

While the data availability of high spatial or temporal resolution remote sensing images continues to increase, it remains in a severely inadequate stage of efficient monitoring using both high spatial and temporal resolution observations (Jiao et al., 2021). Optical remote sensing images have proven to be efficient for water environment monitoring (Cao et al., 2020; Song et al., 2020; Wang et al., 2020). However, there are two limitations associated with previous research. First, critical observations of water changes are limited because of frequent cloudy weather. Especially in tropical and subtropical regions, the availability of optical satellites is insufficient owing to the presence of clouds and rain. Statistics on the widely used Sentinel-2 images indicate that the satellite's usability rate in South China is less than 20% (Figure 6). Second, water change monitoring and driving factor analysis require imaging time covering the entire distribution season, which limits the effectiveness of many applications, for example, seasonal differences in the impact of non-point source pollution on water quality (Lai et al., 2020; Netzer et al., 2019).

Figure 6

Exceptionally low optical image usability rate in South China. According to statistics, for example, the usability rate of Sentinel-2 is less than 20%.

Notably, these two problems are not unique problems to water environment monitoring but are the key issues that urgently need to be resolved in the entire field of remote sensing big data (Jiao et al., 2021). Advancing data fusion methods, such as the combined use of SAR and optical imagery, provide a roadmap for addressing these two methods and improving the effectiveness of water environment monitoring. For example, Fang and Zaitchik (2021) integrated optical satellites (Landsat, MODIS) and SAR images (Sentinel-1) to monitor the changes in wetland water bodies in the Qinghai-Tibet Plateau and their climate response mechanisms. Similarly, Cigna and Tapete (2021) synthetically used In-SAR (ENVISAT and Sentinel-1) and optical satellites (Landsat, Sentinel-2) to explore the impact of groundwater exploitation on groundwater reserves and land subsidence in Mexico's Aguascalientes Valley.

5.3 Low Accuracy of Water Quality Estimation Models Resulting From Complex Water Composition and Insufficient Atmospheric Correction Methods for Waters

The radiance from the water received by the remote sensing sensors accounts for less than 10% of the total radiant energy, most of which is atmospheric radiation (Dornhofer & Oppelt, 2016). Meanwhile, the inherent contradiction is extremely significant between the small range of water spectra and the large range of water quality parameter concentrations (Zhou et al., 2009). Therefore, this requires us to correspond the limited radiant energy with water quality parameters of multiple orders of magnitude. Generally, the Chla concentration and SPM concentration of inland waters are usually within two and three orders of magnitude (Chen et al., 2011; Wang et al., 2017; Zhao et al., 2020). Furthermore, three factors, such as the different sources of optical substances in water, coupling characteristics between water components, and similar attributes of characteristic spectra, make the optical characteristics of waters extremely complex (Zhou et al., 2009). Thus, there is an urgent need to decouple and amplify the optical characteristics of different water bodies and eliminate the absorption and scattering of atmospheric components, such as atmospheric molecules, aerosols, and cloud particles, through atmospheric correction. These preprocesses are the key prerequisites for the high-precision estimation of water quality.

Although great efforts have been made in hyperspectral analysis, radiation transmission simulations, and atmospheric-correction model development, the complex spectral characteristics of water's IOPs result in a lack of high-precision and universal decoupling atmospheric-correction methods that can be applied to various waters. In addition, there are few atmospheric correction methods specifically for inland water bodies. The quantitative estimation of water quality usually uses atmospheric correction methods that were specially developed for terra or open ocean, such as Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH, Tan et al., 2021), “6S” (Second Simulation of Satellite Signal in the Solar Spectrum, Zhu et al., 2020), and ACOLITE (Prakash et al., 2021). These methods are dark pixel atmospheric correction methods that assume that the water reflectance of the NIR band is 0. The scattering of inland water with high suspended solids and Chla concentration, however, is usually significantly greater than 0 in the near-infrared band. As a result, the aerosol optical thickness estimation is too high, which further leads to the failure of the terra atmospheric-correction method and the low estimation accuracy of water quality. Aiming at the improvement of atmospheric correction methods for waters and the great demands for the quantitative estimation of water quality, many researchers have launched a comparative study of atmospheric correction methods. For example, Song et al. (2020) developed and validated regression models for lake clarity in China with different atmospheric correction methods (including calibrated top-of-atmosphere reflectance, Landsat 8 Surface Reflectance Science Product of GEE, Rayleigh-corrected reflectance, and “Dark Spectrum Fitting” atmospheric correction methods) using the Landsat OLI reflectance product. The transparency monitoring results of 40,800 large lakes (>8 ha) in China from 2013 to 2018 show that the atmospheric correction method that eliminates Rayleigh scattering and aerosol scattering has the highest accuracy of the inversion model (N = 1,016, R² = 0.75, and RMSE = 63 cm). In future, more proprietary or universal methods for atmospheric corrections should be proposed and developed by strengthening the relevant research on the coupling mechanism and decoupling methods for water bodies and atmospheric aerosols to reduce the influence of the radiation field, solar flares, underwater reflections, and the basin's underlying surface.

6 Recommendations and Future Directions

6.1 Using Cloud Computing and Emerging Sensors/Platforms to Monitor Water Changes in Intensive Time Series

Effective water environment observations usually require the monitoring of intensive time series with high spatial resolution (e.g., flash floods and agricultural disasters during typhoons). With the emergence of a large number of multi-sensor/platform remote sensing data, water change monitoring has advanced rapidly. In the past half century, there have been a number of major international efforts to address both water extraction and spatio-temporal evolution analysis (Cretaux et al., 2011; Gorelick et al., 2017; Huang et al., 2018; Ma et al., 2015). Likewise, leveraging the Gravity Recovery and Climate Experiment (GRACE) provides a good opportunity for advancing intensive-time groundwater environment monitoring (Feng et al., 2013; Long et al., 2013; Wang et al., 2011). The primary limitation identified in these studies is the contradiction between massive and repetitive image preprocessing and real-time requirements for processing and knowledge discovery (Amani et al., 2020; Ma et al., 2015; Sudmanns et al., 2020). Image downloading, radiometric calibration, atmospheric correction, and other preprocessing account for half of the total time in remote sensing applications. The high-performance computing paradigm and cloud computing platforms are two efficient methods for storing, accessing, and analyzing datasets (Liang et al., 2018). Compared with high-performance computing paradigm, cloud computing platforms have attracted more attention because of its powerful servers and economical and friendly usage. Figure 7 presents an example of the advanced cloud computing platform and its four operating modes: Infrastructure-as-a-Service (IaaS), Platform-as-a-Service (PaaS), Software-as-a-Service (SaaS), and Data-as-a-Service (DaaS). Currently, several cloud computing platforms have handled intensive remote sensing data, such as GEE(Moore & Hansen, 2011), Amazon Web Services (Ma et al., 2015), Microsoft Azure (Amani et al., 2020), International Business Machines (Tamiminia et al., 2020), and Pixel Information Expert Engine (PIE-Engine, https://engine.piesat.cn/). Due to Amani et al. (2020), Ma et al. (2015), and Sudmanns et al. (2020) have detailed the overview, introduction, and applications of various cloud platforms, we only consider the most commonly used GEE as an example to discuss the intensive time application of cloud computing platforms and remote sensing big data for water environment monitoring in this section.

Figure 7

An example of advanced cloud computing platform and its four operating modes, including Infrastructure-as-a-Service, Platform-as-a-Service, Software-as-a-Service, and Data-as-a-Service.

The massively, publicly, and freely available datasets (DaaS) in GEE, as well as its excellent computing performance (SaaS), enable the accurate monitoring of water resources and its environment with sufficient spatiotemporal resolutions (Amani et al., 2020). For example, Landsat TM/ETM+/OLI/OLI2 (1972-present), Sentinel-2 MSI(2015-present), and MODIS (2000-present) in GEE have been widely used for sea surface temperature and salinity (Medina-Lopez & Urena-Fuentes, 2019), surface water dynamics monitoring (Ji et al., 2018; Wang et al., 2020; Zhou et al., 2019), lake bathymetry and long-term mapping (Chen et al., 2017; Murphy et al., 2018; Sagawa et al., 2019; Traganos et al., 2018), suspended sediments (Cao et al., 2019; Griffin et al., 2018), and wetland health assessment (Lin et al., 2018). Furthermore, GEE has a well-configured operating environment, algorithms, and software (SaaS), so users can use it without configuration (Gorelick et al., 2017). Remote sensing cloud platforms, such as GEE, have built a good PaaS architecture for academic ecology. At present, thousands of researchers have shared their research results with peers, enterprise, and management departments through GEE, that is, providing ready-to-use data products such as JRC's data sets of global waters (Pekel et al., 2016). Overall, the advantage of GEE is that it allows remote sensing users to return to the discovery of geoscience knowledge, rather than focusing on downloading and processing a large amount of data. With the rapid development of remote sensing big data and cloud computing platforms, water environment monitoring driven by intensive time series of fusing multi-source remote sensing data based on GEE will provide an effective roadmap for water storage estimations, water pollution monitoring, and climate change monitoring.

The article number of publications related to GEE and water environment was calculated according to the screening criteria of ([TS = 'Google Earth Engine'] or [TS = 'GEE'] and [AK = 'Water']) refined date from 2010-01-01 to 2021-10-28 in Web of Sciences. Subsequently, 64 papers which discussed unrelated topics (e.g., gross ecosystem exchange and 3D Google Map) and 30 patents were discarded. Finally, 1,302 papers were calculated (Figure 8) and their abstract's word cloud was illustrated in Figure 9. Figure 8 indicated that the number of GEE publications on the water environment has rapidly increase since 2017, which is consistent with another study (Amani et al., 2020). Google Earth, Earth Engine, Mapping, Monitoring, Time Series, Land cover, Dynamic, and Analysis were the mostly used words in the abstracts.

Figure 8

The article number of publications related to Google Earth Engine and water environment.

Figure 9

Word cloud of the abstracts from the publications related to Google Earth Engine and water environment.

6.2 Establishing More Models Based on Ensemble Machine Learning to Collaboratively Estimate Multiple Water Constitutes

The advantages of machine learning in fitting the nonlinear relationship between water reflectance and water constitutes provide opportunities for the simulation of the water radiation transmission process and the collaborative estimation of multiple water constituents. After calibration, the model can quickly estimates multiple water constituents. In the past decades, benefiting from the rapid development of computer performance, SVR, ANN, RFR, XGBoost, GA, and PLS have been widely used to estimate water constituents (Cao et al., 2020; Song et al., 2013). However, a single machine learning algorithm has some shortcomings, including overfitting, high-dimensionality, slow convergence, and ease of falling into local extreme values in estimating the water constitutes (Cheng & Han, 2016). Thus, deep learning methods, coupling machine learning models, and ensemble machine learning models have gained more attention in the field of remote sensing big data (Ahmed, 2018; Bai et al., 2021; Song et al., 2012, 2014). For example, the GA-ANN model alleviates the overfitting problem of the classic gradient descent method ANN and obtains the global best solution of nonlinear confrontation in water quality estimation by combining the excellent random learning algorithm GA (Ahmed, 2018; Chen et al., 2021). Notably, the replicate samples by Song et al. (2013) and Song et al. (2014) demonstrated that the average error of the Chla concentration in the laboratory measurements exceeds 8.00%, and the estimation effect of machine learning is greatly affected by the calibrating samples, so the calibrating samples should not only reduce the error, but also try to cover the possible range of values. In the future, multi-parameter water quality estimation, which leverages a large amount of measured data and improved machine learning algorithms, will become a vital research direction in the application of remote sensing big data for water environment monitoring.

6.3 Exploring More Remote Sensing Estimation Models That Couple Physics and Causality

The new generation of Earth observation and remote sensing big data provides opportunities for monitoring water environment characteristics through data fusion, data assimilation, and model development (Smith et al., 2019). Two aspects may need to be strengthened urgently. First, more and broader analysis of IOPs should be enhanced to interpret the physical mechanism of data-driven methods. Data-driven methods represented by empirical statistics and machine learning have the advantages of high simulation accuracy, fewer optical parameters required, and simple algorithms. However, the physical mechanism of data-driven methods is often difficult to explain; and it is difficult to combine the IOPs to discuss simulation errors, which limits the accuracy and generalization ability of the model. In the future, IOPs suitable for complex water bodies such as estuaries, rivers, and reservoirs should be screened out for model optimization and improvement to enrich the physical meaning of data-driven methods. Second, the data assimilation study of continuous water quality monitoring needs to be strengthened. In fact, a single data source and model is difficult to perform fine-grained, high-precision, omni-directional and uninterrupted monitoring of aquatic systems. Based on considering the spatio-temporal distribution of the data and the errors of the observation field and the background field, it will be a useful attempt to integrate the new water environment observation data into the dynamic operation process of the numerical simulation model through data assimilation. For example, data assimilation algorithms (optimal interpolation, variational assimilation, ensemble Kalman filtering, etc.) are commonly integrated with different water environment information to optimize the estimation accuracy of water quality models.

6.4 Identifying More Elements in Water Environment Assessment by Remote Sensing

Given the complexity of the water environment, especially for inland waters, the characteristics of the water environment require interdisciplinary expertise to explore various direct and indirect influencing factors (Jiao et al., 2021). Although some water environment characteristics (optically active parameters) have been widely studied through remote sensing estimation and numerical simulations in the past decades, more features (non-optically active parameters) remain under-explored or even missing (Sagan et al., 2020). Indeed, many studies have focused on the surface water volume and three water color constituents, while estimating the concentrations of more water quality parameters (for example, ammonia nitrogen) are arousing a greater actual demand and becoming more significant. Fortunately, the sensor's band is specially set for the atmospheric window of water radiation, remote estimation models based on hyperspectral (or high-resolution) and machine learning provide opportunities to improve the quantitative estimation accuracy of water quality. When it is associated with economic activities or climate change, it can not only estimate water pollution quantitatively, but also provide scientific and technological support for the treatment of water pollution (McCabe et al., 2017). Therefore, integrating research considering the key factors of water color, water quality, and water volume is urgently needed to promote the transformation ability of science research into actionable responses in the future.

6.5 Developing New Governance Models to Meet the Widespread Applications of Remote Sensing Big Data in Water Environment

Widespread applications of remote sensing big data for water environment monitoring require new governance models. In addition to the Satellite Industry Association (https://sia.org/), many other organizations, including Geoscience and Remote Sensing Society, International Ocean Color Coordinating Group, USGS Water Science School (https://www.usgs.gov/special-topic/water-science-school), European Space Agency (http://www.esa.int/), and Group on Earth Observations (www.earthobservations.org), and IEEE International Geoscience and Remote Sensing Symposium, will address various aspects of water environment applications using remote sensing big data. These organizations can play a helpful role in promoting “the last mile of the application of remote sensing big data” to the water environment and help clarify their vision. They can work with the academic community to provide helpful assistance in the formulation of rigorous scientific standards for remote sensing big data of water environment, especially the documentation and interpretation of uncertainties (Goodchild et al., 2012). Enterprises also provide ethical space and behavioral standards in remote sensing big data for water environment monitoring. Notably, private sectors and volunteers from open sources are currently crucial groups for innovation (McCabe et al., 2017). Private enterprises, such as Google, have competitive advantages, which provide massive data sources and analysis platforms for academia, nongovernmental organizations, and governments through GEE to increase a significant degree of cooperation (Goodchild et al., 2012). Therefore, more similar cooperation between the private sector, academia, non-governmental organizations, and governments for remote sensing big data of water environment may be equally successful.

7 Conclusions

Water dynamics and water quality estimation are fundamental but challenging issues for studying ecological, environmental, and hydrological processes. Water extraction and the quantitative estimation of water quality parameters using remote sensing big data provide effective ways to observe water dynamics and estimate water quality owing to their effective and continuous ability to monitor Earth's surface at multiple scales. During the last decades, considerable efforts have been made to develop various methods for water extraction and the quantitative estimation of water quality in different types of water, including oceans, lakes, reservoirs, and rivers. In this paper, a systematic review of the recent progress in these two fields is presented by comparing a series of water extraction and water quality estimation methods popular in recent literature. For water extraction, we roughly divided the methods into three categories, namely threshold-based methods, water indices, and machine learning-based methods, which have been thoroughly reviewed. For the quantitative estimation of water quality, empirical models, semi-empirical/semi-analytical models, and machine learning-based models are summarized separately. In addition, we reviewed the remote sensing images data sources and model evaluation metrics.

Undoubtedly, these image data sources, water extraction methods, the quantitative estimation methods of water quality, and their evaluation metrics are so limited and cannot meet the great demands of remote sensing big data for water environment monitoring. These shortcomings include: the data gap caused by massive heterogeneous data, insufficient monitoring of water environment in intensive long-term series caused by insufficient available satellite images brought by cloud and rain weather, low accuracy of water quality estimation models resulting from complex water composition and insufficient atmospheric correction methods for water bodies.

Fortunately, we are witnessing the upcoming technological leaps of remote sensing big data, such as data fusion, data simulation, cloud computing, and machine learning. Data-driven approaches based on remote sensing big data, machine learning and cloud computing provide the promising applications for intensive long-term water dynamics and water quality estimation. Physical-driven methods based on radiative transfer equation and data assimilation also provide potential solutions to estimate non-optical matters in long-term and large scales, without limited by the lack of image availability caused by clouds and rain. By reviewing these developments, we do have drawn five promising research directions to optimize the existing framework and methods. These works include: using cloud computing and emerging sensors/platforms to monitor water changes in intensive time series, establishing more models based on ensemble machine learning algorithms to estimate multiple water features collaboratively, exploring more retrieval models that couple physics and causality, identifying the missing elements (e.g., non-optically active parameters) in water environment assessments, and developing new governance models to meet the widespread applications of remote sensing of water environment. This review will be of great benefit to researchers, practitioners, and management departments in the theoretical exploration and innovative application of remote sensing big data for water environment monitoring.

Acknowledgments

This research was financially supported by Natural Science Key Fund Grant of Guangdong Province, China (2018B030311059), Science and Technology Plan Project of Guangdong Province (2018B030320002, 2019A050506001),Natural Science Fund Grant of Guangdong Province, China (2021A1515012579, 2020A1515011068), Key Special Project for Introduced Talents Team of Southern Marine Science and Engineering Guangdong Laboratory (Guangzhou; GML2019ZD0301), and the Water Conservancy Science and Technology Innovation Project of Guangdong Province (publicity no. 2021-21). Authors appreciate the constructive suggestions from anonymous reviewers and editors that help improve this paper.

Conflict of Interest

The authors declare no conflicts of interest relevant to this study.

Open Research

Data Availability Statement

The authors comply with AGU's data policy. Accesses to all the data sets are as follows: The statistics data used for calculating the number of satellites in this study are available at the Union of Concerned Scientists Satellite Database via the link of https://www.ucsusa.org/resources/satellite-database with free access (Union of Concerned Scientists, 2021). Sentinel-2 MSI imagery can be downloaded freely from the Copernicus Open Access Hub (https://scihub.copernicus.eu/dhus/; ESA, 2015).