Application of Projection Learning to the Detection of Urban Areas in SPOT Satellite Images

: We introduce a novel learning algorithm for neural networks, with the major feature of being rapid when compared to classical learning algorithms, offering misclassification rates of 5% and less after only a few iterations, i.e. 20-30 seconds of learning, depending on the task, if a suitable preprocessing has been done. The algorithm is based on considering a neural network as a base in function space, base onto which the function to be learned is projected. We thus call our algorithm projection learning. We present the algorithm, show the application to the detection of inhabited areas in satellite images, discuss the various preprocessors used, compare to other approaches used, and outline further directions of research Key-words: Learning in computer vision, segmentation and perceptual grouping, neural networks, texture analysis, pixel-based classification, low-level processing, feedforward networks, satellite image analysis (R'esum'e : tsvp) weigl@sophia.inria.fr giraudon@soph..