Using Neural Networks in Classifying Satellite images

Abstract

The goal of image classification is to classify all pixels making up multi-spectral images into several land cover classes. The spectral reflectance characteristics of different features manifest different combinations of digital values (spectral signatures). Based on these spectral signatures, a new output image is created having a specific number of categories or clusters. The two general types of classification schemes are the supervised and the unsupervised classification. In this study a neural network classifier, based on the back-propagation algorithm, has been developed and implemented using C++. The classifier is able to detect and differentiate between different land-use features based on their spectral signatures. Classification accuracy is one of the most important aspects of any classification system. This sh1dy showed that the accuracy of a multi-layer network depends on relevant subjects such as: the input feature preprocessing, the order of weights set updating, and the network complexity.

The performance of the neural network classifier has been studied considering the following items:

1. The enhancement of the raw satellite data as preprocessing, which improves the learning process and reduces the learning time. 2. The order of learning samples, which also reduces the learning time. A new proposed method to interleave the samples of different classes gave fast and stable learning

performance especially with higher values of learning rate and momentum. 3. It was proved that, using one hidden layer with a number of hidden neurons less than the number of output classes is enough for doing the classification for any multi-spectral data sets. 4. A modification of back-propagation algorithm has been elaborated by adding lateral connections between the hidden neurons. It has been shown that, this method increases the accuracy of classification in our case study. 5. Both neural networks, with and without lateral connections are compared with that of statistical method (max-likelihood). The neural network with lateral connections gives higher accuracy than neural network without lateral connections, while the statistical method gives the lowest accuracy.

The classification process using the above mentioned modifications has been applied on three different satellite images as follows: TM data covering El Mahala El Kobra area using four spectral bands, TM data covering Siwa area using five spectral bands, and Spot XS data covering Kalyob area using three spectral bands.