STARFM-ML: A novel machine learning driven approach to fuse remote sensing data

There has been an increased interest in combining data from satellites which provide observations at a ne spatial resolution but have extended revisit times, with observations from satellites which provide data at a coarse spatial resolution albeit at a high temporal frequency. Currently the Spatial and Temporal Adaptive Reectance Fusion Model (STARFM) and its subsequent extension for heterogenous landscapes, the Enhanced STARFM (ESTARFM), are the most widely used techniques to fuse such satellite platforms. These algorithms, however, do not take into account the effect of sensor-specific retrieval errors while combining data. They are also unsupervised techniques and thus their application consists of ad-hoc choices of important parameters that govern the overall accuracy of the algorithm. More importantly, these fusion algorithms do not have a framework to incorporate state-of-the-art Machine Learning (ML) algorithms in the fusion process. To account for the above limitations, we propose a novel algorithm called STARFM-ML (ESTARFM-ML) which utilizes the original STARFM (ESTARFM) algorithm in a ML driven state-space stochastic modeling framework. The proposed model is applied to combine daily Evapotranspiration data from satellites ECOSTRESS and MODIS for agricultural sites in the lower Brazos Basin, Texas for the year 2020. When compared with daily Eddy-Covariance measurements, we nd that the proposed approach outperforms the original STARFM/ESTARFM across different landcovers. The proposed scheme is general and can be utilized to combine other environmental variables.