A Pioneering DelugeNet Model with Optimization for Enhanced Urban Flood Detection and Analysis

Hydrological monitoring in urban areas is quite vital in understanding the water flow characteristics and evaluating the propensity of floods for storm water management and flood prevention. Herein, the proposed deep learning model is called DelugeNet which is developed with the aim of refining the extraction of features from flood images and the detection of objects from these images known to play a critical role in improving flood management. With better image features extracted from convolutional layers and other image processing techniques used in DelugeNet, flood related features are captured very well owing to advanced image patterns accompanied by flood phenomena. Upon furthering the knowledge, one of the novel contributions of this work is the deployment of the Elevated Billiard-Inspired Algorithm (EBIA), as a novel optimization that is invoked for calculating the learnable parameter of the deep learning model. Using the EBIA for DelugeNet, it is possible to fine-tune the parameters to achieve better results and increase the reliability in flood detection. A comparison of the accuracy, precision and recall curve proved the suggested approach outperformed other techniques using PASCAL-VOC dataset. The experiments yield clearly better results of DelugeNet over these models, and more significantly, in the detection of crucial flood features of the objects of different classes. For each of the different classes of objects, performances were outstanding with the model attaining: Car-Roof (96.58% accuracy, 96.25% precision, 96.48% recall), Car-Pipe (97.48% accuracy, 97.69% precision, 97.12% recall), Per-Leg (98.58% accuracy, 98.23% precision, 98.14% recall), for mention of others.