Penalty Driven Training Sample Refinement Technique for Hyperspectral Images Classification Using Ant Colony Optimization

Training samples play an important role in learning of a classifier. In order to achieve better spectral information hyperspectral imaging sensors capture reflectance in smaller bandwidth than multispectral sensors. Smaller bandwidth causes lower Signal to Noise Ratio (SNR). Therefore, spatial resolution is kept low in hyperspectral images to improve SNR, but some of the pixels can still be erroneous. If selected in training data, these pixels can cause faulty training which leads to the misclassification. Ant Colony Optimization (ACO) is used to remove erroneous training samples to achieve better accuracy of classification. Available literature suggest reward based methods for selecting more accurate pixels. These methods are not useful if training sample size is already small, as very limited pixels remain available for validation which reduces the efficiency of reward based techniques. In this article, cases with limited number of training samples have been considered for experiments. Results show an improvement of 5-7% in accuracy when reducing the training sample size to 10-30%.