Flood and Non-Flood Image Classification using Deep Ensemble Learning

Floods are one of the most frequent natural disasters, often resulting in widespread devastation. Identifying floods accurately is crucial for disaster management as it helps to locate areas requiring urgent assistance and streamline post-flood evacuation processes. Recently, deep learning models, such as Convolutional Neural Networks (CNN), have become predominant for image classification tasks, as well as flood classification problems. Deep ensemble techniques,i.e. combining several deep learning architectures, are still quite new in many fields and have not been studied extensively despite showing promising results in flood classification. In this research, we develop an ensemble deep learning framework that utilizes eight state-of-the-art CNN architectures, namely MobileNet V2, ResNet 50, VGG 16, DenseNet 201, Inception V3, EfficientNet B5, NasNet Large, and Xception. The aim is to address the gap of deep ensemble learning in flood classification and provide a more effective approach to identifying potential flooding scenarios from a wide range of visual datasets. We utilize FloodNet and flood area segmentation datasets to train, test, and validate our models. In the testing phase, our ensemble model outperforms several individual benchmark models, achieving a training accuracy of 98.9% and a test accuracy of 97.4%. Our proposed methodology will predict floods and conduct early assessments of affected areas efficiently.