Flood Detection and Segmentation Using Deep Learning Models

A flood occurs when water overflows onto normally dry land, often caused by a variety of factors starting from natural events or human actions, resulting in significant losses affecting both human lives and economic stability. This project aims to introduce a flood detection and model by utilizing various deep learning methods. This high clarity proves pivotal for precise flood detection, particularly in segmentation urban locales where buildings and infrastructure might obscure floodwaters. Typically, ordinary media images become accessible immediately following a flood event, unlike satellite images, which may necessitate processing time. This immediacy is crucial for prompt emergency response actions and decision-making processes. The augmentation process applied to non-flooded images encompasses various techniques aimed at enriching the diversity and resilience of the training dataset. Techniques such as image data augmentation are being implemented to improve the model's effectiveness and resilience. The CNN model with attention mechanism demonstrates exceptional performance in accurately categorizing flood images and segmenting flood regions, while the U-Net with a MobileNet base exhibits promising potential. The model achieves impressive scores across various metrics, including accuracy, precision, recall, and Cohen's kappa score, with values of 0.98, 0.97, 0.98, and 0.96 respectively. Additionally, for the segmentation task, the U-Net with MobileNet base algorithm is employed, achieving an accuracy of 0.90, with Mean IOU and Dice Coefficient reaching 0.84 and 0.82 respectively. The CNN model incorporating an attention mechanism exhibits exceptional performance in accurately detecting and segmenting flood regions, while the U-Net with a MobileNet base highlights its potential for flood segmentation. Bayesian Optimization was used to fine tune the hyperparameters and accuracy improved from 90.33% to 91.01%.