Abstract
Severe flood events in India highlight the urgent need for efficient flood management and reliable forecasting systems. One of the biggest obstacles to improving the efficiency of flood monitoring systems is the lack of trustworthy data during flood events. Over the past ten years, computer vision-based methods have become a promising answer for flood monitoring due to recent advancements in information technology. In order to effectively interpret image information and enable meaningful study of flood-affected regions, these approaches mainly rely on robust image segmentation techniques. It is critical for disaster management, particularly in flood forecasting, where precise waterbody detection is essential. However, distinguishing water from visually similar elements such as rooftops, land, and various shades of brown water remains challenging under varying environmental conditions. Traditional methods suffer from misclassification and over-segmentation, affecting prediction accuracy. To address these limitations, we propose a novel superpixel-based segmentation method enhanced with an adaptive erosion technique. Superpixel segmentation effectively groups similar pixels, simplifying image analysis and interpretation, while erosion refines boundaries by removing irrelevant pixel clusters, improving clarity. The final segmentation output is created by applying RGB thresholding to identify water pixels, refining the result using binary erosion, then superimposing the refined mask onto the original colour image. Our method achieves a 1.3% improvement in Jaccard Index, a 3.3% improvement in Recall, a 72.3% enhancement in Boundary F1 Score, and a twofold reduction in computational runtime compared to the SLIC superpixel method, making it a robust tool for flood pre- diction applications.
Keywords: Adaptive Erosion, Computer Vision, Flood Forecasting, Image Segmentation, RGB Thresholding, Superpixel.