Analysis of satellite imagery using a neutral network based terrain classifier

We present a novel method of detecting changes, such as erosion or deforestation, from time sequential pairs of remote images. After preprocessing the images and obtaining a difference image, we use a neural network-based system to adaptively threshold the difference image and resolve areas of pixel intensity with a terrain classifier which combines information in the original images. The result is that we detect precisely the types of changes in which we are interested, without being “distracted” by changes due to noise or natural within-terrain variability of pixel intensity