Post-flood disaster damaged houses classification based on dual-view image fusion and Concentration-Based Attention Module

Flood disasters inflict immense devastation upon buildings, and the post-disaster assessment of housing damage levels is of paramount importance in safeguarding resident lives. Conventional assessment methods entail a significant expenditure of human resources, finances, and time. Toward this, a Dual-View Convolutional Neural Network (DV-CNN) model is proposed in this study. The ResNet-50 model is adopted as its backbone, and the transfer learning techniques and the Concentration-Based Attention Module (CBAM) are incorporated into the model to enhance the efficiency and generalization capabilities of the training model. This model can integrate distinct indicators of interior and exterior damage in post-disaster houses to collectively determine the damage levels classification of the houses. Subsequently, this AI model for house damage level classification is validated by training and testing the data of rural house damage caused by the flood triggered by the “July 20 Heavy rainstorm in Zhengzhou”. The analysis of the model’s hyperparameters indicates that the optimal predictive performance is achieved when the learning rate is set to 0.005, the batch size is 16, and the number of epochs is 50. The results of a comparative analysis among the DV-CNN, ResNet-50, ResNet-101, MobileNet-v2, VggNet, and GoogleNet models indicate that the proposed DV-CNN model achieves the highest accuracy of 92.5% in predicting the damage levels of post-flood affected houses. Finally, a visual analysis of the features associated with house risk levels provides a clear understanding of the classification mechanism and the accuracy of the model. Experiments demonstrate that the deep CNN recognition model based on dual-view exhibits greater reliability and generalizability, providing a valuable reference for classification models of post-flood damage levels in rural houses.