A high-accuracy ensemble method of convolutional neural networks for damaged building detection on post hurricane satellite images

After hurricane disasters, rapid damage assessment is critical to ensure an effective response. However, the current method for building damage assessment requires intense time and human resources. The emerging high-resolution optic satellite, computer vision, and machine learning techniques could help improve the efficiency of damage assessment. In this work, an ensemble method of convolutional neural network was proposed for high accuracy of automatic damaged building detection of post-hurricane satellite images. The customized convolution neural network, deep learning architectures of random initialization and transfer learning were trained separately. Then a model average ensemble was introduced to avoid overfitting and further increase the detection accuracy. The hurricane Harvey dataset consisting of 23000 images has been used in this paper for training and testing. The experiment results illustrated that the ensemble model can achieve very high accuracy above 99% on both balance and unbalanced test datasets, outperforming all the previously published results.