Sensitivity of the subspace method for land cover classification

The quality of a supervised classification map depends on the quality of the ground reference data and the classification method used. However, training samples for agriculture landscapes are often mixed with noise. Therefore, the classification of agriculture regions using remotely sensed data requires the use of classification methods with good generalization capabilities. In this study, the performance of the subspace method in land cover classification of a complex cropping mix area is explored. Landsat-5 thematic mapper (TM) data were used to classify 12 different land cover classes in the study area, located between Tianjin and Tangshan cities in northern China. We compared the classification maps obtained using the subspace method with those obtained using the self-organizing map neural network (SOM) and maximum likelihood classification (MLC) methods. The results of this comparative study confirm that the subspace method performed better than both the SOM and MLC methods. Furthermore, a comparison of the sensitivity of these methods to the reduction in the training sample size shows that the subspace method has a lower sensitivity to variations in the number of training pixels used than the other two methods. Our results demonstrate the ability of the subspace method to distinguish between different crop types over a large area. Moreover, the subspace method is less sensitive to small training sample sizes than the other two methods.