Automated Detection of Ship Tracks in Multispectral Satellite Data

Under a variety of conditions, effluents from ship stacks can enhance reflectively of shallow layer clouds in such a way that the presence of ships can be detected in daytime satellite images. Multispecial observations from AVHRR abroad NOAA polar orbiting satellites have been used to identify ship tracks through enhanced cloud reflectivity at 3.7 microns, and to a lesser extent 0.63 microns. Nonetheless, some difficulties remain associated primarily with the high variability of cloud reflectivity which results from large variations in liquid water amount and droplet sizes at small spatial scales. Further, the age of the tracks can lead to signatures that are weak and, hence, can pose problems for automated detection, even though the tracks are apparent by visual inspection. In this proposal we address the problem of developing an automated ship track detection system. To do so, we propose to perform the following: 1) optimally enhance satellite images using multispectral signals; 2) determine features that best characterize higher reflectivity and curvilinearity of ship tracks; 3) apply rule-based and cluster analysis techniques to reduce the data stream to a limited number of subscences with potential tracks; 4) apply state-of-the-art neural net and statistical discriminant analysis methods as final detection filters; 5)assess detection success and error rates; 6) develop an automated prototype system design. The algorithms used here are ones that have exhibited success on this or similar problems. We will adopt the approach of applying a sequence of tests or detection filters to optimize the benefits of each and to minimize detection error rates of both types.