Stratified and adaptive regression for NDVI-based land surface temperature downscaling over agricultural fields

Improving the spatial resolution of thermal imagery is essential for agricultural field management, especially in developing countries where fields are fragmented and heterogeneous. The existing downscale land surface temperature (LST) products are a promising solution while awaiting for further advances in satellite-based thermal sensors. In this study, a linear regression-based approach was proposed to downscale LST from 1 km to 10 m spatial resolution. The proposed method is hereafter referred to as Stratified and Adaptive Regression for Land Surface Temperature Downscaling (STAR-LST). Taking advantage of the linearity between LST and NDVI under homogeneous conditions, multiple linear regression models were first derived for different regions in the (LST, NDVI) feature space at coarse resolution. These regions were defined by partitioning the triangular form of the (LST, NDVI) feature space into sub-triangles, sharing a common top vertex. Each model is then applied to the corresponding NDVI range, allowing thus to derive LST at finer resolution. LST estimates were validated against in-situ measurements collected over olive trees, Kernza, wheat, and barley. Good results were obtained with RMSE values ranging from 2.79 to 4.49 °C for STAR-LST. This presents an improvement of an average of 33 % compared to the two classical methods DisTrad and TsHARP (RMSE in the range 4.18–6.83 °C), which showed similar accuracy. In addition, STAR-LST enabled daily LST estimation whenever coarse-resolution data were available, using interpolated NDVI to generate a high spatio-temporal LST product suitable for managing agricultural fields, particularly the small size fields.