A Random Forest Approach for Generating Daily Surface Reflectance Time Series Integrating MODIS and Sentinel Data

A novel approach is presented to generate a series of continuous global surface reflectances, preserving sensitivity to disturbances such as forest fires. Daily Random Forest models estimate missing reflectance values in MODIS-TERRA bands NIR and SWIR using data from the corresponding MODIS-AQUA bands, Sentinel-3 Synergy bands (Oa17, S3N, and S5N), and auxiliary data (DEM, slope, aspect, hillshade, land cover, sun zenith and azimuthal angles, sensor angles). The output of the models for each pair of bands is assembled for the final reflectance estimates. Data integration involves MODIS image preprocessing for alignment with Synergy products. The reflectance estimation exhibits an average RMSE lower than 0.04 in the NIR and SWIR bands. The reflectance time series obtained allow effective identification of fires and show low variability (0.20-0.35) in stable unburned canopies. The method is computationally efficient and readily applicable to other years.