From Pixels to Precision: Deep Semantic Segmentation of Water Bodies in Remote Sensing Imagery

Abstract

Accurate and precise identification of water bodies from remote sensing imagery plays an important role in various fields like environmental monitoring, resource management, and disaster assessment. This study presents a comprehensive exploration of two distinct yet interconnected approaches: semantic segmentation and fractional mapping, aimed at enhancing the precision and applicability of water body analysis. The primary purpose of this research is to use RGB image datasets instead of multi-spectral water bodies datasets. The research focuses on using robust and computationally efficient U-Net architecture. This choice is rooted in its ability to achieve efficient semantic segmentation results while demanding significantly fewer training resources. In addition to semantic segmentation, this study extends its focus to fractional mapping, an approach aimed at quantifying the distribution and extent of water bodies through continuous values. However, the unavailability of suitable datasets for fractional map ping poses a challenge. To solve this, a custom dataset of 100 lakes is created using satellite imagery from planet.com. The research uses various preprocessing techniques and loss functions, ensuring optimal results. For fractional mapping of water bodies, we used the same U-Net model which was adopted for semantic segmentation of water bodies. The research used different stan dard loss functions and introduced novel hybrid loss functions, tailored to segmentation. The trained model’s performance is evaluated on unseen datasets, with semantic seg mentation inference conducted on test data, and fractional mapping inference carried out on the Namal Lake dataset. For semantic segmentation, the U-Net model shows the best results across IoU-Focal-LCD loss in terms of validation IoU coefficient which is 79% and Dice-Focal loss shows the best value for validation Dice coefficient which is 83%. For fractional mapping, the U-Net model with Huber loss shows 85% validation and 95% training accuracy.