Deep Transfer Learning for Automated Artillery Crater Classification

Sani Aji, Poom Kumam, Punnarai Siricharoen, Ali Maina Bukar, Mohammed Sani Adamu

Abstract

Analysis of artillery craters is an indispensable tool due to its importance in military investigation and other precautionary purposes. It is used to verify suspected locations of hostile fire, detect the presence of enemy ammunition as well as tracing the direction from which weapons are fired. Traditionally, crater analysis is done manually. Not only that the manual process consumes time, but also, they are prone to errors and sometimes, result in loss of crater information. Hence, the need for automation. Recently, deep neural networks have recorded great breakthroughs and achieved promising results in different applications. In this paper, we propose the use of deep convolution neural networks (DCNNs) for the automatic classification of artillery craters. Specifically, due to the small size of training data, transfer learning is employed. We used satellite images to fine tune five pretrained DCNN models, VGG16, VGG19, ResNet50, MobileNet and EfficientNetB0. Interestingly, all these pretrained models achieved impressive classification accuracy in classifying craters from other environmental potholes. To our knowledge, this is the first paper that consider low-angle fuze craters. Our proposed approach is automatic, and can assist military and other human right investigation in automatic crater analysis instead of the manual ways.