Detecção de elementos antrópicos em imagens aéreas da floresta amazônica

Abstract

During environmental crimes patrolling, the response time is a very important component for the success of the missions. Generally, infractions occur in remote and hard-access places, characteristics that hinder both the patrolling as well the action of environmental protection agents. To increase the approaches’ success rate and reduce the risk of human lives, unmanned aerial vehicles (UAVs) can be used to cover large areas of forest in a short time without being perceived by offenders, allowing the patrolling organs responsible for these areas to plan and act more efficiently in the repression of such crimes. The new problem generated by this approach is the huge amount of data generated during these missions, which often includes hours of video. The manual inspection of all this material in searching for anthropic elements is very tiring and error-prone. This work presents a evaluation of image segmentation techniques, inspections of features to be extracted, followed by a supervised classification of those segments for anthropic element detection in amazon’s rain forest aerial images. Besides making publicly available a dataset with more than 3,000 images and 10,000 segments labeled accordingly, this work investigates different strategies for anthropic elements classification. The experiments obtained a consistency error rate inferior to 8% in image segmentation and a precision above 94% on target objects classification through one-class classifiers ensemble, using One-class SVM and REPTree algorithms.