Detection of ship tracks in AVHRR cloud imagery with neural networks

An artificial neural network (ANN) based hybrid detection system is proposed and developed for fast, accurate, and automatically processing multispectral satellite cloud imagery to identify ship tracks. Imagery observed at both 11.0, 3.7 and 0.67 μm by multispectral Advanced Very High Resolution Radiometer (AVHRR) NOAA polar-orbiting satellite is used in this development. A data base containing 5 different categories of ship tracks based on their contrast and edge information have been created. After image enhancement and morphology based pre-selection, a back propagation trained and feed forward neural network classifier is successfully developed to extract, learn, and identify desired ship track information over an area containing portion of track segments. Preliminary results show that well-formed ship tracks can be successfully identified at low false alarm rate. The effects due to various extraneous backgrounds are also investigated