Refining Semantic Segmentation of Flood Images Using Edge Sharpening and CNN

Abstract

Post-disaster impact analysis is an important component in supporting mitigation planning, emergency response, and evidence-based decision-making after flood events. Visual data, such as flood images, can be used to identify affected areas and analyze environmental conditions through semantic segmentation. Semantic segmentation is a pixel-level classification process that assigns each pixel in an image to a specific object class. However, flood images collected from real-world conditions often have low visual quality, unclear object boundaries, and complex backgrounds, which may reduce the quality of segmentation results. This study proposes an edge-sharpening-based preprocessing approach combined with a Convolutional Neural Network (CNN) model to improve semantic segmentation performance on flood images. The proposed method applies unsharp masking to enhance edge and contour information before the images are processed by the CNN model. The experiments were conducted using flood and non-flood image datasets and compared with the ENet baseline and a modified ENet model. The evaluation was performed using visual comparison and quantitative metrics, including precision, recall, F1-score, accuracy, and mean Intersection over Union (mIoU). The results show that the proposed method achieved the best performance on the flood image dataset, with 98% precision, 98% recall, 98% F1-score, 97% accuracy, and 54% mIoU. These results outperform the comparative CNN models and indicate that edge sharpening can improve object boundary representation, particularly for flood images with blurred or low-quality visual characteristics.