Improved remote-sensing land cover identification using parameterized feedforward neural networks

This article proposes methods of parameterizing the inputs of a feedforward neural network (FFNN) that classifies the land cover (LC) in remote-sensing (RS) images in order to speed up the classification process while maintaining high classification accuracy. FFNN training was optimized via two parameters, the learning time (LT) and the number of neurons (NN), by decorrelating the LC data in the RS image using either a discrete wavelet transform (DWT) or independent component analysis (ICA), although ICA can only be applied to a multiband image. The RS images used in this work have also been the focus of several previous attempts at LC classification using various methods. They consisted of a 4.6-m airborne resolution image with HH polarization that was acquired by a synthetic-aperture radar (SAR) and a 15-m multispectral resolution image acquired by the Terra satellite’s Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). Corresponding field-truth data was used to validate the FFNN classifiers examined in this work. High resolution gives high accuracy information, so that the LC correlation increases, in this case. Classification results for the ASTER image showed that the NN of the FFNN classifier was reduced by more than half when the classifier was parameterized using a DWT, and by three-quarters when the classifier was parameterized using ICA. Results for the SAR image indicated that the NN of the FFNN classifier was halved when the classifier was parameterized using a DWT. Parameterization also reduced the LT of the classifier. The classification accuracy was assessed using a confusion matrix. The fast parameterized FFNN classifiers presented strong classification performance characteristics, similar to those of the original FFNN classifier, with overall accuracies that always exceeded 0.75 and sometimes reached 1. Subsequent work should focus on optimizing the FFNN further by automating two steps: (1) image decomposition using ICA or DWT and (2) FFNN classification.