Flood Area Segmentation using Semantic Segmentation-based Deep Learning Models

Abstract

Floods are natural disasters that frequently occur in different areas of Pakistan and cause several damages including human lives, infrastructure, and material losses. The unavailability of bench mark dataset of the flood data is the primary constraint in improving the flood monitoring systems. With state-of-the-art computer vision and deep learning techniques, flood detection, segmentation, and recognition systems have gained much attention. In this paper, a deep learning based DeepLabV3 architecture has been applied to segment the flood areas from the images. A ResNet-50 model has been used as a backbone network for extracting the features. The model has been trained and evaluated on the flood area segmentation dataset. The performance of the flood area recognition model has been assessed using the mean F1-score and Jaccard index. This research study also compares the performance of DeepLabV3 architecture with other backbone models including VGG-16 and DenseNet. Furthermore, the performance of UNet architecture has been evaluated and compared with DeepLabV3. Based on the experimental results, the DeepLabV3 architecture with ResNet-50 as a backbone model achieved the best segmentation results than the other models.