An Automatic Extraction of Features from Water Bodies by Using Contrast Vision Transformer

In recent days, for strengthening the climate change adaption water features are utilizing as the most crucial environmental element. The integration of Remote Sensing (RS) with Artificial Intelligence (AI) is becoming the major sought-after approach for extracting the features of water information. In existing, the Faster Regional Convolution Neural Network (FR_CNN) is utilized for the water bodies' extraction. However, the existing FR_CNN model struggled with computational complexity. To overcome this issue, Contrast Vision Transformer (CoVT) is proposed to extract features from water bodies. The transformer with self-attention mechanism reduces the computational complexity by eliminating regional proposal networks. Initially, the satellite images are taken from sentinel-2 dataset and further processed into preprocessing. In preprocessing Gaussian Median Filter (GMF) and Atmospheric Correction (AC) are utilized to reduce the noise and atmospheric effects from images. After that, proposed CoVt extract features from preprocessed images and fed into projection head to generate projected features. Finally, the projected features are further processed with prediction head to predict the satellite image features of water bodies. From the results, the proposed CoVt provided better results when compared to existing FR_CNN by providing results such as accuracy (99.4%), precision (99.1%), recall (98.5%) and F1-score (98.7%) respectively.