Application of neural networks for identification of sea ice coverage and movements from satellite imagery

Visible and infrared (0.67, 0.8, 3.7, 11.0, and 12.0 μm) imagery from the Advanced Very High Resolution Radiometer (AVHRR) aboard NOAA´s operational meteorological satellite provides a high resolution (1 km×1 km) measurement and unique signatures for the identification of sea ice coverage and movements in polar regions. A backpropagation trained artificial neural network (BP ANN) algorithm is developed and applied for the classification of several sea surface conditions, including open water, grease, young ice, and multi-year thick ice coverage from the multispectral information of AVHRR satellite imagery. The trained BP ANN classifier is applied to images covering same region at successive days in order to investigate the changing of the physical status and the movement of ice blocks. An associated neural network based on Hopfield network architecture is investigated to determine the movement of ice coverage at two consecutive time series measurements from classified images. This network is derived from the cross-correlation analysis through minimization of the least mean square error between two images. Displacement and motion at each pixel can be obtained from the output values of the neural network