Object-based image analysis approach for vessel detection on optical and radar images

Commercial satellites for Earth observation can integrate conventional positioning and tracking systems for monitoring legal and illegal activities by sea, in order to effectively detect and prevent events threatening human life and environment. This study describes an object-oriented approach to detect vessels combining high- and medium-resolution optical and radar images. Once detected, the algorithm estimates their position, length, and heading and assigns a speed range. Tests are done using WorldView-2, QuickBird, GeoEye-1, Sentinel-2A, COSMO-SkyMed, and Sentinel-1 data imaged in several test sites including China, Australia, Italy, Hong Kong, and the western Mediterranean Sea. Validation of results with data from the automatic identification system shows that the estimates for length and heading have R2  =  0.85 and R2  =  0.92, respectively. Tests for evaluating speed from Sentinel-2 time-lag image displacement show encouraging results, with 70% of estimates’ residuals within ±2  m  /  s. Finally, our method is compared to the state-of-the-art search for unidentified maritime object (SUMO), provided by the European Commission’s Joint Research Centre. Finally, our method is compared to the state-of-the-art SUMO. Tests with Sentinel-1 data show similar results in terms of correct detections. Nevertheless, our method returns a smaller number of false alarms compared to SUMO.