A hybrid methodology based on Neural Network and Change detection approaches and using Sentinel-1/Sentinel-2 for soil moisture estimation

Soil moisture is an essential parameter for a better understanding of water processes in the soil-vegetation-atmosphere interface. In this context, passive and active microwave remote sensing have enabled the development of various increasingly operational approaches, in particular for low spatial resolution products. Synthetic aperture radar (SAR) is particularly suitable for monitoring water content at fine spatial resolutions of the order of 1 km spatial resolution. Since the launch of Sentinel-1 in 2014, numerous methodologies have been proposed for estimating fine spatial resolutions soil moisture, especially in agricultural areas. Two approaches are often considered in the inversion of SAR signals: approaches based on machine learning methodologies, such as neural networks trained on scattering models, or approaches based on change detection, essentially validated on low spatial resolution products using scatterometers. In this study, we propose a hybrid approach combining both the neural networks and change detection approaches. The methodology was applied to Sentinel-1 and Sentinel-2 using numerous predictors; Vertical-Vertical (VV) polarization radar signal, incidence angle, Normalized Difference Vegetation Index (NDVI) optical index, VH/VV ratio, etc.

This hybrid approach is tested on the database of the international soil moisture network (ISMN) with moisture networks covering different climatic contexts. Results are very encouraging with 10% improvement in the accuracy of soil moisture estimates compared to the use one of each approach individually (Neural Network or change detection).