Identificação de incêndios florestais utilizando segmentação de imagens e aprendizado de máquina

Abstract

This work proposes the use of different data preprocessing and deep learning techniques for image analysis and forest fire detection. The images used for training originate from two different databases with variation in time, weather season and positioning. For training, we chose to employ supervised learning algorithms and probabilistic classifiers, totaling three training sources with parameter variations and different complementary pre-processing techniques, such as color perception and quartiles. The main evaluation metric refers to accuracy and the rate of true positives and false negatives, essential for this application, as it is an identification and alert system. Processing and training time values are also considered. The results obtained were superior to the state-of-the-art for identifying forest fires, with accuracies greater than 99.6% using the Random Forest technique.