Object detection in high resolution optical image based on deep learning technique

Abstract

Deep learning-based object detection in high resolution optical imagery is an active research frontier. Dataset for object detection is fundamentally important to drive deep learning models. Compared to datasets in computer vision, there is a lack of datasets for object detection in the remote sensing field. This work produced a high-quality dataset for object detection based on the high-resolution remote sensing images of the TripleSAT. The spatial resolution of the TripleSAT images is 0.8 ​m and the dataset for object detection includes 3 categories: wind turbine, airplane and oil storage tank. 3583 TripleSAT images are used, in which 7320 instances are manually annotated, to compose a TripleSAT dataset. To test the performance of the TripleSAT dataset, a state-of-the-art deep learning model, the RetinaNet was used with the PyTorch framework. Our results show that the mean average precision of the RetinaNet for wind turbines, airplanes and oil storage tanks are 93.47%, 92.8% and 96.55%, respectively. The results demonstrate the capability of the TripleSAT satellite images in producing datasets for deep learning models. Dataset made from single remote sensing sensor can achieve high accuracy and fast detection. The proposed dataset, including wind turbine, oil storage tank and airplane is meaningful supplement for data on hazard-affected bodies.