Land-use classification of gray-scale aerial images using probabilistic neural networks

With the advancement in remote sensing methods that can provide high-resolution data, it has become important to develop classification methodologies to exploit these technologies. The objective of this study was to develop a probabilistic neural network (PNN) based technique for the classification of gray-scale aerial images into various types of land use, especially for rural areas where agriculture is important. We defined specific land-use classes including city, water, forest, and various types of agricultural field areas. Gray-scale aerial images with a 6.5 m resolution for nine counties in Georgia were obtained from the Georgia Geographic Information Clearing House. Three approaches were used for the preparation of the inputs to the PNN, including histograms of the pixel intensities, textural parameters extracted from the image, and matrices of pixels for spatial information. Twelve hundred images were used for model development, and an additional 300 images were used for model evaluation. The best PNN was based on textural parameters, with an overall accuracy of 92% for the evaluation data set, while the histogram approach had an accuracy of 90% and the spatial approach had an accuracy of 66%. Several combinations of all three approaches were also evaluated, but they did not provide an improvement in accuracy. Further work is needed for practical implementation of the best classification procedures.