Machine Learning for Moisture Retrieval Using Landsat and SAOCOM Images

Soil moisture (SM) estimation is crucial for environmental, hydrological, and agricultural applications. The satellite remote sensing technology allows monitoring SM in a consistent and economically viable manner. Thus, the Argentine Microwave Observation Satellite (SAOCOM) mission uses a Synthetic Aperture Radar (SAR) system to monitor SM with high spatial resolutions. Despite the advancement of SAOCOM, this mission was recently launched in 2018, therefore it covers a short period of SM data, i.e. approximately five years. On the contrary, the Landsat program has been monitoring the earth's surface for longer than SAOCOM. The Pampas region is an important productive zone of Argentina, mainly characterized by agricultural land, so monitoring SM in this region results of paramount importance for agricultural productivity. This study evaluated different machine learning algorithms to reproduce the SAOCOM SM maps using Landsat images in the Cordoba province. Specifically, three machine learning methods were explored: Random Forest (RF), Multilayer Perceptron (MLP), and a deep learning model based on the U-Net architecture. The modelled SM maps were compared with SAOCOM images and their soundness was evaluated using the Mean Squared Error (MSE) and Pearson correlation (r) statistics. Our experiments show that the U-Net model yields the best performance in comparison with SAOCOM SM maps with a correlation of 0.94. Indeed, U-Net shows a MSE of about 16%, while RF and MLP present errors of about 29% and 28%, respectively. These results suggest that the U-Net architecture with Landsat images can capture the SAOCOM spatial variability and allow monitoring SM with high spatial resolution in an extensive temporal series.