Feature Selection For Tree Species Identification In Very High Resolution Satellite Images

The aim of this study was to provide an effective feature selection for tree species classifiers in mixed-species boreal forest, from a very high resolution optical satellite image. The 35 input features were the 5 input spectral bands (multispectral and panchromatic channels), 9 contextual features derived from the panchromatic channel and 21 segment-wise features computed at three segment sizes around the treetop locations. A variable ranking was first performed to evaluate the relevance of each feature. Then sequential forward selection was carried out using k-nearest neighbors (kNN) and Linear Discriminant Analysis classifiers. The results suggested that a reasonable feature set would contain 6 to 10 features, mostly from input bands and contextual features. On such a feature set, the best kNN classifier (k=5) returned classification accuracies of 76% for pine and spruce and 88% for decidous trees, with RMS errors between 1.4% and 3.5% and few mixing with the 4 non-tree classes.