Damgage Images Assessment by Means of Convolutional Neural Network Combined with Transfer learning

Abstract:
It’s always a concern for human beings when facing natural disasters, which includes hurricane, an especially damaging one. Damage assessment is one significant way for government and rescue team to draw up a rapid plan, which can be time-consuming if conducted pure manually. With the rapid development of deep learning model and convolutional neuron network, these methods are widely applied to image classification areas. To find out those damaged places in a short period, this paper proposed a deep learning model using transfer leaning based on VGG19 and MobileNet. By the training process on the images collected by satellites, the model can classify images to damaged and undamaged labels with high accuracy. Some data augmentation methods are also applied, including rotation and flip. The models based on VGG19 and MobileNet respectively all reach an accuracy over than 94% on two test sets, while the ensembled model reached an accuracy over 96% on two test sets, which implies the effectiveness of the model.