IDENTIFICATION OF A FOREST SPECIES OF SOCIAL-ENVIRONMENTAL INTEREST BASED ON MACHINE LEARNING

Abstract

A new deep learning method capable of mapping unique Mauricia flexuosa palm species, known as Buriti in aerial RGB imagery was proposed. Buriti is an important palm tree for communities and fauna, in addition to being an indicator of path areas, and its mapping is important. The first session of this dissertation presents a brief report on the world flora and legislation pertaining to the studied species. A second session presents a new method based on the Convulational Neural Network (CNN) that makes it possible to identify and geolocate a species, in addition to comparing its performance with other object detection networks. They were manually labeled a total of 5.334 palm trees in a set of 1,394 ortoimage oatches, returning a mean absolute error (MAE) of 0.75 trees and F score of 86.9%. The results are better than Faster-RCNN and RetinaNet methods. Concluding that the proposed method is efficient to deal with high density and forest complexity environment, being able to map and locate as species of Mauricia flexuosa with hight accuracy.