Análise de estratégias para identificação de copas de árvores em áreas arbóreas utilizando imagens de Aeronaves Remotamente Pilotadas

Abstract

The monitoring of forest areas is an essential tool for environmental conservation, either by controlling the suppression of native vegetation or by monitoring the restoration processes, with constant improvements in the strategies and techniques used in their execution. Therefore, it is necessary to know the structure of these areas in order to have a better understanding of their dynamics and distribution. In this scenario, the use of geoprocessing tools, especially through the technique of segmentation of the treetops, appears as essential tool to study vegetation in detail, assisting in the study of the structure of the forest and its conservation. However, its applicability is often limited by the quality of the images available for analysis and, sometimes, by the techniques used for data interpretation. Remotely piloted aircraft (RPAs) are becoming an excellent alternative in terms of data quality, allowing the acquisition of high-resolution images at low cost. In addition, with easier access to machine learning techniques, it is possible to use neural network models to process data in an innovative and efficient way. In this work, images obtained through RPA were used to evaluate different techniques for segmentation of tree crowns, aiming to find the most suitable to identify and count tree individuals. The following segmentation methods were evaluated: mode filter, multiresolution, watershed and segmentation with neural networks for deep learning. The images were obtained in areas with different plant physiognomies (Pinus silviculture and riparian forest), in order to assess their impact on the results. The results were compared using a set of validation data (terrestrial truth) generated from independent samples of these same images, to observe the accuracy of the techniques. The results showed that the polygons from the mode filter do not correspond to the crowns of the trees, possibly due to the intense homogenization of the images, which emphasizes its use as an intermediate segmentation tool. The multiresolution and watershed segmentations, on the other hand, showed good potential to detect treetops in the Pinus area, however, its parameters must be adjusted according to the target area, which makes it difficult to use and requires the algorithm to be executed multiple times. Ultimately, the segmentation generated by the deep learning model (U-Net) had problems connecting the edges of the treetops and showed a result without individual polygons for each tree. However, with the labeling used in this project, the tool showed good potential for detect gaps within the forest. In all cases, results were better in Silviculture compared to the area of Riparian Forest, a situation that may be related to the difference in the density of the treetops between these regions. Finally, additional experiments are necessary, using other sensors (such as LiDAR) and methods, to check if there is an improvement in the quality of the segmentation.