Semantic Segmentation for Ships Detection from Satellite Imagery

Ships detection, as well as other object detection, and localization tasks in satellite images are the central problems in the field where remote sensing and computer vision coalesce. They are commonly used in different areas like environment monitoring, fishery management, logistics, insurance and many others. This paper provides an approach based on the Convolutional Neural Networks (CNN) as the main algorithm/instrument for detecting ships in optical satellite images of different spatial resolution. For achieving the best performance, we divided the problem into stages, which gave a possibility to control the quality of intermediate outcomes. The proposed method contains two parts: 1) building a classifier based on XCeption, 2) using baseline Unet model with Resnet18 as encoder for exact segmentation which allow us to achieve accuracy of more than 84%.