Automated detection and geolocation of teak trees based on deep learning from very-high-resolution rgb uav imagery

Abstract

The control, supervision and monitoring of trees, both in terms of their location and height, is of vital importance in teak plantations (Tectona grandis L.f.). This allows planning silvicultural work related to thinning, replanting, irrigation and phytosanitary treatments, estimating production and harvesting. Knowing the number of plants, their spatial location, and estimating their height is crucial for managing large tree plantations and determining their carbon sequestration capacity to contribute to climate change mitigation. In this work, three different size versions of the anchor-free and single-stage detector YOLOX deep learning network, pre-trained on the COCO dataset, were specifically trained for automatic localization of teak trees in large and dense plantations from UAV imagery. This study used two teak plantations located in Ecuador. "La Marina" plantation (456 ha) served as the training and validation area, while "La Selena" (195 ha), was reserved for testing and accuracy evaluation, constituting a true holdout dataset to assess the generalization capabilities of the developed model. Very high-resolution RGB images were taken in both plantations using a Phantom 4 drone at a flight altitud of 120 m above ground. Regarding the results obtained, they showed that the "small" version of the YOLOX deep learning network performed significantly better than the other two versions tested ("medium" and "large" size YOLOX), presenting notably good average metrics of Precision (94.74%), Recall (82.49%) and F1-score (87.91%). In this sense, the trained model proved to be a suitable solution to address complex visual recognition challenges in very high-resolution UAV images.