作者： Taskin Kavzoglu , Paul M. Mather
DOI:
关键词: Mahalanobis distance 、 Machine learning 、 Time delay neural network 、 Neural gas 、 Pattern recognition 、 Feature selection 、 Artificial intelligence 、 Probabilistic neural network 、 Computer science 、 Deep learning 、 Recurrent neural network 、 Bhattacharyya distance
摘要: Feature selection is an important issue, especially for classification problems where artificial neural networks are involved. It known that using large number of inputs can make the network overspecific and require significantly longer time to learn characteristics training data. Such over-specificity also reduces generalisation capabilities a network, so may fail classify new data outside range Although feature methods have been used in remote sensing studies many years, their use context has not fully investigated. This paper sets out some results investigation techniques, specifically separability indices, problem determining optimum structure terms achieved accuracy. For this purpose, including divergence, transformed Bhattacharyya distance Jeffries-Matusita distance, Mahalanobis classifier (MDC) based on two accuracy measures employed determine best eight-band combination 24 band multitemporal dataset. Two search procedures, sequential forward genetic algorithm, combinations as evaluation functions.