Finding Ship Tracks with Machine Learning Algorithms in Remote Sensing Images

Ship-tracks are an iconic demonstration of aerosol-cloud interactions. They have been observed in satellite images for more than 60 years. Many studies have resulted from analyses of remote sensing, in-situ, and modeling results. Much have been learnt from these efforts. However, manually identifying ship tracks from satellite imagery is extremely costly, which prevents large, global scale study of ship tracks. Here we present an algorithm that is trained with about 1000 manual samples and can automatically identify ship tracks with high fidelity. The input for this algorithm only requires two IR channels that are widely available on many legacy, ongoing and upcoming sensors, which renders the algorithm applicable for a wide range of data. The algorithm not only identifies ship tracks that are picked up by humans but also finds those that are easily missed by human eyes from a first glance. We demonstrate the power of this algorithm through applying it to one year worth of MODIS data. The results are very encouraging. Global ship track maps have excellent agreement with those shipping lane density maps from shipping industry, which suggests our algorithm is working properly. Future work will be to improve the algorithm in terms of efficiency and apply it to a range of research topics.