Texture Based Supervised Learning for Crater-Like Structures Recognition Using ALOS/PALSAR Images

摘要

Meteorite impacts participated in the formation of the Solar System and continue to modify the planetary surfaces, originating a structure present in all of them, the craters. Terrestrial craters are abundant, geological and biological significant structures and are related to large mineral ores. The Earth impact record continues to be deciphered, currently 190 terrestrial impact structures have been confirmed, and it is estimated that several hundred remain to be discovered. One of the techniques to detect a crater candidate site is Remote Sensing, however it is a difficult task, due to the large information that must be processed, the lack of discriminant features for crater and non-crater regions and appropriated methods to recognize them. We propose an approach to identify meteorite impact structures, based on textural features of ALOS/PALSAR grayscale radar images, using supervised automatic learning. For this, the quotient of HV and HH polarimetric bands of these images was calculated. The resulting images were segmented by global thresholding to generate two sets of training samples: structure type and regions type of craters and non-craters, with them different kinds of classifiers (Bayesian, Fuzzy, Genetics, Bagging, and Boost) were trained, getting accuracy between 81 to 99% for craters identification.