Satellite data is crucial for monitoring soil conditions, with applications in agriculture and environmental management. This study assesses soil moisture in a semi-arid region of Tunisia using Sentinel-1 satellite imagery and CNN-based classifiers developed for time-series data classification. The training database was validated with Sentinel-2 imagery and ground-truth data to enhance classification accuracy.

The study area, in central Tunisia's Kairouan governorate, spans the eastern Tunisian Atlas (9°30′E to 10°15′E, 35°N to 35°45′N). Measurements were taken during a 2019 field campaign. The region's land use includes cereals, orchards, olive groves, bare soils, fallows, vegetables, urban areas, and dams. Sentinel-1 and Sentinel-2 images from June to October 2019 were downloaded from the Copernicus platform. Sentinel-1 data was preprocessed using ESA’s SNAP software, involving terrain correction, noise reduction, and radiometric calibration, and then segmented in QGIS using GPS field reference data. Statistical analysis in R revealed correlations between VH backscattering, land uses, and NDVI with rainfall. Sentinel-1 (VH polarization) and Sentinel-2 NDVI images informed sampling hypotheses for CNN training, considering soil moisture and biomass variations.

The CNN classifiers, implemented in MATLAB R2018a, were evaluated through cross-validation, selecting models that minimized errors and maximized accuracy. The CNN achieved the best results with a learning time of 503 seconds, accuracy of 99.36%, and a loss of 6.76%. The training phase used 9x9 sampling windows, 15,000 samples (2/3 for training, 1/3 for testing), 5 classes, 50 filters (5x5), Maxpooling (window size = 4, stride = 2), and the activation function 'gradient descent with momentum.' The learning rate was 0.005, with 1 fully connected layer, 30 epochs, and 50 iterations. The CNN outperformed the Random Forest (RF) method applied to Sentinel-2 data in handling data complexity such as moisture, biodiversity, and biomass.