A physics-informed machine learning approach to estimate surface soil moisture

We propose Physics Informed Machine Learning (PIML) algorithms to estimate surface soil moisture from Sentinel-1/2 satellite images based on Artificial Neural Networks (ANN). We have used Improved Integral Equation Model (I2EM) to simulate the radar images backscatter in VV polarisation. In addition, we selected a set of different polarisations, i.e.; (VH, VH/VV, VH-VV), incidence angle, Normalised Difference Vegetation Index (NDVI), and topography as input features to map surface soil moisture. We have used two different approaches to predict soil moisture using PIML. In the first approach, we developed an observation bias in which we selected the difference of backscatter value at each pixel in VV polarisation from satellite and derived from theoretical model derived as one of the input features. Our second approach is based on learning bias, in which we modified the loss function with the help of the I2EM model. Our result shows the learning bias PIML outperforms the observation bias PIML with R = 0.94, RMSE = 0.019 m³/m³, and bias = -0.03. We have also compared the performance with the standalone benchmark algorithms. We observed the learning bias PIML emerged as the most accurate model to estimate the surface soil moisture. The proposed approach is a step forward in estimating accurate surface soil moisture at high spatial resolution from remote sensing images.