Three methods (fuzzy partition method, stepwise regression analysis and principal component analysis) were used to select meaningful texture features for discriminating forest cover types. The initial texture set was extracted from the wavelet sub-images. Feature selection was based on all texture features of four sub-images combined. Recognition of forest cover types was accomplished by the neural network of learning vector quantization. The performance of these techniques was evaluated using a case study area at Whitecourt, Alberta, Canada. The selection procedure seemed to be adequate to extract meaningful texture features to help discriminate forest cover types, because the classification accuracy of the selected feature sets was improved. In addition, the optimization process can be considered as an efficient one, since the number of features was reduced to about 24.5-66.8% of the total 208 features using the three selection methods.