Abstract

River segmentation from Remote Sensing Imagery (RSI) has significant research value and practical applications for monitoring river changes, comprehending patterns in river water levels, flood detection, agricultural planning, and environmental monitoring. Therefore, monitoring river areas and water bodies is essential. This paper proposes the deep-learning approaches based on a decreased Convolutional Layer U-Net based (dConvLU-Net) method to perform an efficient segmentation of river and land from RSI containing inference of non-river information such as bridges, shadows, and roads. The results of the experiments show how well these models work compared with other semantic segmentation models in many aspects of river water segmentation. The FCN-based method takes less execution time and the least computational cost but the mean Pixel Accuracy(mPA) and Mean Intersect over Union (mIOU ) are also less. U-Net performs better mPA and mIoU despite their increased computational costs and execution time. However, the dConvLU-Net method performs most effectively regarding execution time, computational cost, mPA, and mIoU. The results of the proposed dConvLU-Net method show that the river segmentation from RSI is fast and accurate.

Authors

Miral J Patel¹, Hasmukh P Koringa², Jay Pandya³, Darshan Tank⁴
Government Engineering College, Rajkot, India^1,2,3, Government Polytechnic, Rajkot, India⁴

Keywords

Deep Learning, Segmentation, RSI, U-Net