Supervised desertification classification using Siamese Variational Autoencoder

Accurate desertification detection is critical for environmental management in arid regions. This paper presents SVAEDesert, a supervised classification framework based on Siamese Variational Autoencoder (SVAE) to improve desertification detection. The proposed framework efficiently extracts features from bi-temporal satellite images using a variational autoencoder (VAE), minimises clustering errors with the Gaussian distribution, and improves the spatial consistency of the extracted features using the Siamese structure with shared weights. A supervised classifier based on a Soft Max layer is trained on these representations to generate desertification maps that distinguish between low, high, and very high-risk areas. Evaluated on satellite data from southern Tunisia, SVAEDesert outperforms state-of-the-art models such as Siamese CNN (SCNN) and Siamese RNN (SRNN) with an accuracy of 97.43%, a precision of 97.42%, an F1-score of 97.39%, a recall of 96.86% and a false positive rate (FPR) of 0.026. These results demonstrate the robustness and effectiveness of the model for reliable desertification monitoring.