Radial basis function(RBF) neural network was applied to determine soil types of hilly and mountainous terrains in Fengdu County of the Three Gorges region in China, the elevation in which ranges between 118.5m and 2000m, combining landsat enhanced thematic mapper plus (ETM+) data and topographic information from a digital elevation model (DEM). We designed a RBF network using newrb P,T,GOAL,SPREAD) function in MATLAB software, in newrb function orthogonal least squares learning algorithm be used to choose Gaussian kernel function centers and the weights of the network. Two sets of training samples were selected for training. One was a set of 3606 training samples; the other was a set of 57905 training samples, also for maximum likelihood classification. Considering training time, we divided these 57905 samples into 3063 small sample areas, so a set of averages of which was selected to input the network for training at last. The classification results with RBF neural network showed that the second training samples set generated 60.3% producer's accuracy, higher than that of the first samples set. But the producer's accuracies of RBF neural network trained by both sets were lower than that of using maximum likelihood classifier with the same training samples, which was 66.6%. On the other hand, the Kappa coefficient of RBF neural network trained by the second training samples set was 0.5587, higher than that of maximum likelihood classifier with the same training samples, which was 0.4919. So, it is indicated that RBF neural network for soil classification is not the best method under limited training samples for more samples more training time consumption to the unacceptable extent.

Date Published: 3 November 2005
PDF: 7 pages
Proc. SPIE 6043, MIPPR 2005: SAR and Multispectral Image Processing, 604327 (3 November 2005); doi: 10.1117/12.654987

Published in SPIE Proceedings Vol. 6043:
MIPPR 2005: SAR and Multispectral Image Processing
Liangpei Zhang; Jianqing Zhang; Mingsheng Liao, Editor(s)