A comparison of different regression models for downscaling Landsat and MODIS land surface temperature images over heterogeneous landscape

Abstract

Remotely sensed high spatial resolution thermal images are required for various applications in natural resource management. At present, availability of high spatial resolution (<200m) thermal images are limited. The temporal resolution of such images is also low. Whereas, coarser spatial resolution (∼1000m) thermal images with high revisiting capability (∼1day) are freely available. To bridge this gap, present study attempts to downscale coarser spatial resolution thermal image to finer spatial resolution using relationships between land surface temperature (LST) and vegetation indices over a heterogeneous landscape of India. Five regression based models namely (i) Disaggregation of Radiometric Temperature (DisTrad), (ii) Temperature Sharpening (TsHARP), (iii) TsHARP with local variant, (iv) Least median square regression downscaling (LMSDS) and (v) Pace regression downscaling (PRDS) are applied to downscale LST of Landsat Thematic Mapper (TM) and Terra MODIS (Moderate Resolution Imaging Spectroradiometer) images. All the five models are first evaluated on Landsat image aggregated to 960m resolution and downscaled to 480m and 240m resolution. The downscale accuracy is achieved using LMSDS and PRDS models at 240m resolution at 0.61°C and 0.75°C respectively. MODIS data downscaled from 1000m to 250m spatial resolution results root mean square error (RMSE) of 1.43°C and 1.62°C for LMSDS and PRDS models, respectively. The LMSDS model is less sensitive to outliers in heterogeneous landscape and provides higher accuracy when compared to other models. Downscaling model is found to be suitable for agricultural and vegetated landscapes up to a spatial resolution of 250m but not applicable to water bodies, dry river bed sand sandy open areas.