Evaluation of LST downscaling algorithms on seasonal thermal data in humid subtropical regions of India

Abstract

Thermal image downscaling algorithms use a unique relationship between land surface temperature (LST) and vegetation indices (e.g. normalized difference vegetation index (NDVI)). The LST–NDVI correlation and regression parameters vary in different seasons depending on land-use practices. Such relationships are dynamic in humid subtropical regions due to inter-seasonal changes in biophysical parameters. The present study evaluates three downscaling algorithms, namely disaggregation of radiometric surface temperature (DisTrad), sharpening thermal imagery (TsHARP), and local model using seasonal (25 February 2010, 14 April 2010, and 26 October 2011) thermal images. The aggregated Landsat LST of 960 m resolution is downscaled to 480, 360, 240, and 120 m using DisTrad, TsHARP, and the local model and validated with aggregated Landsat LSTs of a similar resolution. The results illustrate that the seasonal variability of the LST–NDVI relationship affects the accuracy of the downscaling model. For example, the accuracy of all algorithms is higher for the growing seasons (February and October) unlike the harvesting season (April). The root mean square error of the downscaled LST increases from 480 to 120 m spatial resolution in all seasons. The models are least suitable in water body and dry-river bed sand areas. However, the downscaling accuracy is higher for NDVI > 0.3. The present study is useful to understand the applicability of the downscaling models in seasonally varied landscapes and different NDVI ranges.