Multisource data selection for lithologic classification with artificial neural networks

Mapping the geological lithology in remote regions is a difficult and costly endeavour. Multisource image data such as electro-optical reflectance and microwave backscatter combined with geophysical image data such as gravity, magnetic, gamma ray spectrometry, offer the potential to construct lithologic maps without extensive field surveys. Neural networks are a useful tool for classifying image data, since they bypass the statistical assumptions required by traditional multivariate discriminant functions, and they are much simpler to setup than expert systems. However, the performance of neural networks can vary significantly between individual classes. In this work we examined the effects of the different input image data on network performance. We found that the geophysical data, particularly radiation measurements, which provides subsurface information, were essential for accurate mapping of lithology. Mid-infrared reflectance was the most relevant of the remote sensing measurements. We also found that the learning styles of the various networks varied considerably. In all cases, a complete learning history for each network was required before the network can be trained for optimal performance.