Feature matching from SAR Arctic data using neural networks

Matching sea ice features consists of recognizing the same features of ice in two different images. Feature matching can be used as a tool in different applications, like ice motion tracking, assisting in feature classification, and determining temporal changes of ice features. The authors describe how they use neural networks to compare different features and decide how likely they are to be the same. A backpropagation neural network was trained to determine how closely different features match, considering possible deformations. The input to the network is the contour of each feature. Different input patterns were studied, and a pixel grid representation of the features contour is shown to yield better results than a chain code representation. A database containing ice features from different images was used as an implementation tool for extracting the desired data. The efficiency of using such a database for extracting data from sea ice features is shown and some details are given about its implementation. The effectiveness of the neural network in matching features in different images over time is demonstrated and its applications on tracking ice motion and assisting on features classification is explained