Improving Land Surface Temperature Estimation in Cloud Cover Scenarios Using Graph-Based Propagation

Abstract

Land surface temperature (LST) serves as an important climate variable which is relevant to a number of studies related to energy and water exchanges, vegetation growth and urban heat island effects. Although LST can be derived from satellite observations, these approaches rely on cloud-free acquisitions. This represents a significant obstacle in regions which are prone to cloud cover. In this paper, a graph-based propagation method, referred to as GraphProp, is introduced. This method can accurately obtain LST values which would otherwise have been missing due to cloud cover. To validate this approach, a series of experiments are presented using synthetically obscured Landsat acquisitions. The validation takes place over scenarios ranging from between 10% and 90% cloud cover across six urban locations. In presented experiments, GraphProp recovers missing LST values with a mean absolute error of less than 1.1°C, 1.0°C and 1.8°C in 90% cloud cover scenarios across the studied locations respectively.

Key Points

• Gaps in satellite-derived land surface temperature (LST) measurements caused due to clouds can tackled using graph-based propagation

• The proposed approach, GraphProp, recovers missing LST values more accurately than existing tensor completion methods from literature

• The presented results show the approach to be robust in even highly challenging settings including up to 90% cloud cover

Plain Language Summary

Although it is possible to work out what the ground temperature is from satellite images, when there are clouds which cover part of the image we cannot use existing methods. In parts of the world which are often cloudy, this means you might only occasionally have access to information about the surface temperature. Land surface temperature relates to many physical processes, therefore, we propose a method which aims to provide accurate surface temperature values even when it is cloudy, in order to assist the study of these systems. To do so, the method we propose uses a satellite image of the same region which was captured on a different day when there were not clouds. This earlier image is used to describe the region in a graph structure, where similar areas are grouped together. This representation of the region is then used to work out what the region might have looked like if there were not clouds in the later image. From this, we can apply existing methods to compute the land surface temperature. In our experiments we show that this approach is effective and the recovered temperature values are more accurate than can be recovered by other existing approaches.