Using Neural Networks In Remote Sensing Image Processing

Abstract

Current methods of terrain classification of remotely sensed images are not always accurate enough to be easily incorporated into an automated system for terrain classification and analysis. One of the obstacles encountered has been the development of suitable mathematical models to use for classifying the various types of datasets. Also, most methods do not allow for the inclusion of 'uncertain classes' at various stages in the analysis process, which would allow for reclassification later on as more detailed information is subsequently obtained. Recent research into neural networks at CCRS (the Canada Centre for Remote Sensing) has shown good potential for their classification capabilities with the use of simple raw input data. Furthermore, it has been demonstrated that these promising results have been obtained without the necessity of developing complex and specific mathematical models: neural networks automatically create mathematical solutions to the problems presented to them. This paper addresses further investigation into the application of neural networks at various stages in the classification of a remotely sensed image. We plan to do this in three ways: 1.Providing datasets in better prepared and preprocessed format to the neural network, 2. Providing more appropriate datasets as input to the neural network by including more parameters for class identification, and 3. Attempting to allow for 'uncertaih classes' in the form of dynamic allocation of a new classes, whenever necessary. Neural networks perform their classification tasks better if they have properly preprocessed datasets. This means that the data must be as clean and as clear as possible. To do this the data can be pre-filtered for noise using neural nets. It can also be transformed, or pre-manipulated in such a way that the channels provided for input are more finely tuned for subsequent class identification. Any classification system can make a general classification with a basic representation set, and with the inclusion of more identifying characteristics in the representation set, a more detailed and definitive classification can be achieved. In other words, the more parameters that are made available to the system, the easier it is to identify items in the dataset. Therefore we will experiment with more appropriate datasets by including spatial parameters along with the conventional spectral channels as input to the neural network. Finally, there are always some items in a dataset which cannot be properly identified on the basis of the representation set the system has available to it at the time. Accordingly, we are trying to accommodate this fact with the provision of uncertain classes to be reclassified at a later stage in the analysis process. In conclusion, we hope to be able to illustrate that neural network technology is another useful tool which can be added to the list of conventional tools to improve the accuracy of terrain class identification for inclusion into an automated expert system.