Classifying multispectral data by neural networks

Two methods are tested for improving multispectral neural network classification: (a) new criterion functions and (b) incorporating contextual information. In the first approach, several energy functions for synthesizing neural networks are tested on 2-D synthetic data and on Landsat-4 thematic mapper data. These new energy functions, designed specifically for minimizing misclassification error, in some cases yield significant improvements in classification accuracy over the standard least mean squares energy function. In addition to operating on networks with one output unit per class, a new energy function is tested for binary encoded outputs, which result in smaller network sizes. The thematic mapper data (four bands were used) is classified on a single pixel basis, to provide a starting benchmark against which the contextual approach is compared. For single pixel classification, the best neural network result is 78.7%, compared with 71.7% for a classical nearest neighbor classifier. The 78.7% result also improves on several earlier neural network results on this data. In the contextual approach, several methods are tested, all employing the basic idea of concatenating the spectral values of neighboring pixels to the spectral values of the pixel to be classified. The best result was obtained by including spectral values from the 4 nearest (horizontal and vertical) neighbors, which increased the classification accuracy from 78.7% to 80.4%. Insight is provided into the nature of the classification errors by comparing the ground truth and spectral classification images. Approaches for further improving accuracy are discussed, including feature extraction methods for reducing the dimension of the feature vectors while still retaining contextual information.