Increasing Landsat 5 TM Spatial Resolution to 15 M Using a Super-Resolution Deep Learning Model Trained with Pan-Sharpened Landsat 7 ETM+ DATA

In this work, we aim to recover the information at 15 m spatial resolution from Landsat 5 TM (L5 TM) data with 30 m spatial resolution using a super-resolution deep learning model. The model is designed to predict a pan-sharpened Landsat 7 ETM+ (L7 ETM+) image at 15 m resolution from a L5 TM image at 30 m spatial resolution. For the model training, we used images of L5 TM and L7 ETM+ from the same region and at a time interval of acquisition < 10 days. Our results show that the model achieves to improve the spatial resolution of the L5 TM even with a modest sample for training, constituted only of 4225 couples of L5 TM and L7 ETM+ images of size 128 ×128 pixels from the Landsat tile 216068. We also found that L5 TM emulated reflectance values at 15 m spatial resolution were more comparable to the values and ranges of L7 ETM+ reflectance than the original L5 TM reflectance. With an improved dataset for training, this model could be used to produce a dataset of images spatially and radiometrically harmonized of L5 TM and L7 ETM+ at 15 m spatial resolution.