Utilização de redes neurais convolucionais e imagens obtidas por RPA para o mapeamento de palmeiras na Amazônia Ocidental

Abstract

The palm trees (Arecaceae) are one of the most important resources from the social and economic point of view for the local communities in the Amazon, because they guarantee income and provide resources such as food, raw material for construction, handicrafts, and industry. The complexity of Amazonian forests limits obtaining crucial information for commercial exploitation and management of palm trees, such as density and spatial distribution. We evaluated the performance of the YOLOv4 Convolutional Neural Network in the automatic detection and classification of palm trees in native tropical forests. The study was conducted in a remnant of Open Ombrophylous Forest in southwestern Amazonia. First, an RGB orthophoto was generated from images obtained with an Unmanned Aerial Vehicle. The orthophoto was then subdivided into 960 plots of 37.5 x 37.5 meters. We manually labeled 1098 palm trees identified by photointerpretation belonging to four palm species: Attalea butyracea (Mutis ex L.f.) Wess. Boer, Euterpe precatoria Mart., Iriartea deltoidea Ruiz & Pav and Oenocarpus bataua Mart. Data augmentation was performed to increase the learning ability of the model. It randomly selected 80% of the data for training, 20% for validation. To make the predictions, the Artificial Neural Network for object detection YOLOv4 was used. The method achieved overall average accuracy of 91.08% and the average accuracy for A. butyracea, E. precatoria, I. deltoidea and O. bataua was 92.07% ±2.85%; 96.2%; ±1.48%; 93.83%; ±3.09% and 92.48% ±2.82%, respectively. YOLOv4 is an important tool for mapping palm trees in native forests, serving as a support for forest planning and management.