Abstract
An optimized immune network-based classification (OPTINC) method was developed and adapted for land use / land cover classification. Based on the widely-used artificial immune network (aiNet) model, three major improvements were made: (1) the preservation of the best antibodies for each class from being suppressed; (2) the usage of self-adaptive mutation rates in response to changes in model performance between learning generations; and (3) the integration of genetic algorithm-optimized linear combinations of Euclidean distance and spectral angle mapping distance as affinity measurements. OPTINC was evaluated for two study sites with multi-sensor data. Decision trees, neural networks and aiNet were also tested and compared in terms of classification accuracy, local homogeneity of the classified image, and model sensitivity to sample size. OPTINC outperformed the other models with higher accuracy and much less salt-and-pepper noise in the classification images. OPTINC was relatively less sensitive to training sample size than decision trees and neural networks were.