Digital lithology mapping from airborne geophysical and remote sensing data in the Melville Peninsula, Northern Canada, using a neural network approach

Feedforward neural networks were used to identify bedrock types from satellite images and airborne geophysical data of the Melville Peninsula, Northern Canada. The first experiment, on rock outcrop pixels, demonstrated that more than 90% classification accuracy was achieved with as few as 150 training samples (0.5% of “rock outcrop” pixels). The second experiment consisted of re-building the bedrock geology map based on all pixels (except water pixels) in the area. The results demonstrated that greater than 85 % classification accuracy can be obtained with only 150 training samples (0.17% of all pixels in training area). Tests with more training samples were also performed in both experiments and only a moderately higher accuracy was obtained. In both experiments the geophysical data were the controlling factors in the classification. The Landsat TM data contributed to classification results only for the experiment on rock outcrop pixels.