Deep learning and aerial imagery for macaúba palm identification

Abstract

The objective of this work was to use deep learning and images taken by unmanned aerial vehicles to develop a model to identify the occurrence of macaúba (Acrocomia intumescens) palm trees. The model was trained and tested using data from areas in the southern region of the state of Ceará, Brazil. Later, the tested model was evaluated using data from areas in the Midwestern region of the country. The primary challenge was to distinguish macaúba from other native palm trees, such as babassu (Attalea speciosa). Babassu has spectral similarities and a random distribution, which makes it difficult to identify. Red-green-blue mosaics were cropped into smaller size images and processed using a convolutional neural network deep-learning technique. Identification performance was evaluated using metrics of accuracy, precision, recall, and F1-score. In an area of 1,000 ha, 3,679 macaúba palm trees and 12,410 babassu palm trees were identified, achieving a 93% accuracy. The proposed approach was evaluated in a 4.0 ha site located in the municipality of Batayporã, in the southern region of the state of Mato Grosso do Sul, with an 89% accuracy. The model was able to identify macaúba palm trees occurring in natural areas in the Semiarid and in Midwestern regions of Brazil.