Unsupervised self-training method based on deep learning for soil moisture estimation using synergy of sentinel-1 and sentinel-2 images

Here, we present a novel unsupervised self-training method (USTM) for SM estimation. First, a ML model is trained using the labeled and unlabeled data. Then, the pseudo-labeled data are generated employing the second model by adding a proxy labeled data. Eventually, SM is estimated applying the third model by pseudo-labeled data generated by the second model and unlabeled data. The final SM estimation result is obtained by training the third model. Subsequently, in-situ measurements are performed to validate our method. The final model is an unsupervised learning model. Experiments were carried out at two different sites located in southern Tunisia using Sentinel-1A and Sentinel-2A data. The input data include the backscatter coefficient in two-mode polarization (σ ° VV and σ ° VH), derived from Sentinel-1A, as well as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Infrared Index (NDII) for Sentinel-2A and in-situ data. The USTM method based on (Random Forest (RF)- Convolutional neural network (CNN)-CNN) combination allowed obtaining the best performance and precision rate, compared to other combinations (Artificial Neural Network (ANN)-CNN-CNN) and (eXtreme Gradient Boosting (XGBoost)-CNN-CNN).