MRDenseNet: automated building damage detection model after the earthquake with a combination-based deep feature engineering

Abstract

Earthquakes have caused significant damage wherever they occur throughout the world. It is important to assess the effects of earthquakes rapidly and accurately after they occur. This study presents a new deep featureengineering model for post-disaster building-damage detection. We gathered a new image dataset, and this dataset contains 2,716 building images belonging to three classes: (1) damaged, (2) debris, and (3) non-damaged. We present a deep feature engineering method: a combination-based information fusion model. Three pretrained convolutional neural networks (CNNs) MobileNetV2, ResNet50, and DenseNet201 were used for feature extraction. Based on its components, the image classification model was termed MRDenseNet. A feature vector was extracted from each CNN (thus, there are three main feature vectors in total). Seven feature vectors were constructed using combinations of the feature vectors (i.e., two cubed minus one We eliminated redundant features from each vector using iterative feature selection. The K-nearest neighbor (kNN) and support vector machine (SVM) classifiers were applied to the seven feature vectors to obtain 14 results. Voted results were obtained in the information fusion phase, and the greedy algorithm was used to select the best output. MRDenseNet generates 14 classifier-based outputs and 12 voted outputs. Thus, 26 candidate outputs were generated, and MRDenseNet selected the best-performing output as the final result. Thus, MRDenseNet can be described as a self-organized classification model. The MRDenseNet model attained a 94.15% accuracy for the three-class case. Furthermore, the MRDenseNet reached a 97.72% classification accuracy for detecting damaged buildings.