An Artificial Immune System Approach for Unsupervised Pattern Recognition in Multispectral Remote-Sensing Imagery

This paper presents an improved Artificial Immune System (AIS) approach for unsupervised classification in multispectral remote-sensing imagery. For benchmarking, one has considered several unsupervised nature-inspired intelligent classifiers (AIS, neural, fuzzy) versus statistical ones. We have comparatively evaluated the following pattern recognition techniques: the proposed AIS model; Self-Organizing Map (SOM); Vector Quantization SOM (VQSOM); Fuzzy C-means, and K-means. The considered techniques have been evaluated using both synthetic and real datasets. The real datasets correspond to the LANDSAT 7 ETM+ multispectral image (341 x 343 pixels) taken in June 2000, representing a region of Bucharest, Romania. There have been considered four pattern classes: artificial surfaces, agricultural area, forest, water. One has also evaluated the case of choosing a balanced dataset from the LANDSAT image, with equal number of 800 selected multispectral pixels per class. For the balanced LANDSAT dataset with 3 bands (1, 4, 5), the best experimental correct recognition score is of 93.78% for AIS model followed by the scores of 89.09% for the 5 x 5 neuron SOM model, 83.28% for VQSOM, 84.18% for Fuzzy C-means, and 83.15% for K-means. Key-Words: multispectral remote-sensing imagery, unsupervised pattern recognition, nature-inspired intelligent techniques, Artificial Immune System (AIS), Self Organizing Map (SOM), VQSOM, Fuzzy C-means, K-means.