SPOT Panchromatic Imagery and Neural Networks for Information Extraction in a Complex Mountain Environment
Additional Document Info
High resolution satellite imagery is often required to obtain accurate information about landforms and the terrain. In mountain environments, the magnitude, frequency, and interaction of lithospheric and atmospheric processes cause high topographic and spatial reflectance variability. Consequently, information extraction is difficult and new approaches are required to assess and map complex spatial patterns. The purpose of this research was to evaluate the utility of artificial neural network (ANN) technology for recognizing spatial reflectance variation related to alpine glacier characteristics. Specifically, we wanted to determine if a minimally trained neural network could be used to map the supraglacial characteristics of glaciers on the Nanga Parbat massif in Pakistan. We trained a three-layer feed forward network using the back-propagation learning algorithm to recognize reflectance variations from SPOT Panchromatic data. We compared ANN classification results to a stratified unsupervised classification approach using the ISODATA clustering algorithm. Results indicated that minimal training of a ANN is sufficient to produce accurate information regarding supraglacial characteristics. Overall classification results were relatively high with kappa coefficients ranging from 0.85 - 0.91. Accuracy assessment and comparative visual analysis indicated that ANN performance was superior to the performance of the ISODATA algorithm. Results demontrate that there is significant potential associated with using ANN technology for information extraction and for mapping complex spatial patterns in mountainous terrain. 1999 Taylor & Francis Group, LLC.