Class feature-weighted hyperspectral image classification

This article develops a new approach to hyperspectral image classification (HSIC), called class feature (CF)-weighted HSIC (CFW-HSIC), which extracts features from classes of interest that can be used to improve HSIC. Three types of CFs, intra-CFs (Intra-CFs), inter-CFs (Inter-CFs), and total CFs (TCFs), are particularly designed to calculate probability for each of the classes as its class significance that can be used for classification. The use of such CF-calculated probabilities is threefold. One is to develop an algorithm to automatically allocate a specific training sample size for each of the classes for classification. Another is to use the CF-calculated probabilities as class weights to generalize commonly used overall accuracy (OA) to CF-OA. A third one is to take advantage of CF-calculated probabilities to derive CFW-HSIC, which can effectively deal with classification of unbalanced classes as well as class variability/class separability. Most importantly, CFW-HSIC can perform better than HSIC without using CFs.