Water Resources Identification in Barren Lands Using Deep Learning Techniques

This paper introduces a hybrid deep learning architecture to precisely detect surface water resources in deserted and dry areas in Central India with Sentinel-2 RGB images. Conventional spectral techniques are affected by spectral confusion and this is particularly evident in dry settings where water features appear as shadows or murky soil. To mitigate this, a U-Net architecture based on ResNet-50V2 encoder is suggested, which uses transfer learning to increase its ability to extract features. The model is trained on the Sen-2 LULC dataset of 213,761 images, and a combined Binary Cross-Entropy loss function with Dice loss function to manage high class imbalance (less than 5% water pixels). The experimental results indicate high performance with the highest IoU of 0.92 and a precision of 0.89, recall of 0.91, and Dice coefficient of 0.94. The suggested algorithm is superior to baseline U-Net and is a scalable solution to monitoring water resources in arid regions.