Ship detection using optical satellite imagery is of great significance in many applications such as traffic surveillance, pollution monitoring, etc. So far, a lot of ship detection methods have been developed for images covering open sea, offshore area and harbors. Compared to the ship detection in sea and offshore area, it is more difficult to detect ships in inland river due to several challenges. First of all, many ships in inland river are clustered together and hard to be separated from each other. Secondly, ships lying alongside the pier are very likely to be recognized as part of the pier. Thirdly, ships in inland river is usually smaller than those in the sea. A hierarchical method is proposed to detect the ships in inland river in this paper. The Regions of Interest (ROIs) are firstly extracted based on water–land segmentation using multi-spectral information. Then two kinds of ship candidates are extracted based on the panchromatic band. The isolated ships are detected by analyzing the shape of connected components and the clustered ships are detected by using mixtures multi-scale Deformable Part Models (DPM) and Histogram of Oriented Gradient (HOG). At last, a Back Propagation Neural Network (BPNN) is trained to classify the ship candidates using the multi-spectral bands. The experiments using Quickbird satellite images show that our approach is effective in ship detection and performs particularly well in separating the ships clustered together and staying alongside the pier. © 2019 Elsevier Inc.