Individual tree detection and quantification is a crucial step in ecology, essential for forest management, particularly in conservation units, and automating this process remains a challenge. Remote sensing (RS) can be a powerful ally in this regard, enabling the coverage of large areas. By combining aerial images with Deep Learning (DL) models, it is possible to achieve automated detection and contribute to integrating new technologies into the field of ecology. For this reason, the objective of this study is to analyze effective approaches for tree crown detection based on DL in RGB images obtained through an Unmanned Aerial Vehicle (UAV) in dense ombrophylous forests and Amazonian savannas. Images were selected from the Alter do Chão Environmental Protection Area (APA) and the Tapajós National Forest (FLONA), where RGB images were collected with a Ground Sample Distance (GSD) of 3.26 cm/pixel using an UAV to create 1 km 2 orthomosaics in TIFF format. After preprocessing, the images were subjected to different training configurations in the DeepForest model, varying the number of epochs (10, 15, and 30) and resolution (4 cm, 7 cm, and 10 cm). Among the results obtained, the model achieved an F1 Score of 0.8290 with 30 epochs, compared to 0.7451 with 10 epochs. Using the best resolution (4 cm), F1 Scores of 0.9021 were achieved for the denser savanna and 0.7984 for the FLONA. The results demonstrate that the DeepForest model could be adapted to the specific conditions of the Amazon, one of the most complex biomes in the world. • First use of DeepForest in Amazon Conservation Units for crown quantification. • We introduce a new dataset with images and annotations of tree crowns in the Amazon. • Local data in supervised learning improve tree crown quantification in the Amazon. • The model generalizes well to new images, even under diverse conditions.
Silva et al. (Fri,) studied this question.