Uso de visão computacional para inventariar árvores urbanas na UTFPR campus Londrina

Abstract

This work presents research on the implementation of computer vision techniques in urban tree mapping, using images captured by drones and deep learning models. The main focus was to analyze the predictions and limitations of the DeepForest and YOLOv8 algorithms in the automatic detection and counting of trees on the campus of the Federal Technological University of Paraná - Londrina. The methodological approach included the manual annotation of 177 trees, image preprocessing in blocks of different dimensions and overlaps, elimination of blocks without relevant information, and the application of artificial data augmentation techniques, allowing the training of models under different configurations. Performance was evaluated through traditional computer vision analyses, such as accuracy, recall, F1 Score, and mean accuracy (mAP). The results indicated significant differences: DeepForest exhibited greater sensitivity, with a recall exceeding 60%, but presented a high number of false positives, while YOLOv8 demonstrated superior selectivity, reaching up to 69% accuracy, although with very low recall, resulting in a specific undercount of trees. This discrepancy highlights that neither model, in isolation, managed to achieve an adequate balance between reliability and comprehensiveness, although both showed promise for practical applications. The main limitations identified in the study were the limited number of annotated trees and the limited number of available images, which restricted the diversity of examples for training and validating the models. Even so, the work represents a pioneering initiative at UTFPR - Londrina, demonstrating that the use of artificial intelligence can reduce costs and time in urban tree inventories, implemented as a starting point for future investigations encompassing more extensive databases, more recent architectures, semi-supervised development strategies, and the integration of results into georeferencing systems.