Precision silviculture: use of UAVs and comparison of deep learning models for the identification and segmentation of tree crowns in pine crops
Manuel Pérez-Carrasco, Bruno Karelovic, Roberto Zayas Molina, Rodrigo Saavedra, Pierluigi Cerulo, G. Cabrera-Vives
Abstract
The monitoring of trees is crucial for the management of large areas of forest cultivations, but this process may be costly. However, remotely sensed data offers a solution to automate this process. In this work, we used two neural network methods named You Only Look Once (YOLO) and Mask R-CNN to overcome the challenging tasks of counting, detecting, and segmenting high dimensional Red–Green–Blue (RGB) images taken from unmanned aerial vehicles (UAVs). We present a processing framework, which is suitable to generate accurate predictions for the aforementioned tasks using a reasonable amount of labeled data. We compared our method using forest stands of different ages and densities. For counting, YOLO overestimates 8.5% of the detected trees on average, whereas Mask R-CNN overestimates a 4.7% of the trees. For the detection task, YOLO obtains a precision of 0.72 and a recall of 0.68 on average, while Mask R-CNN obtains a precision of 0.82 and a recall of 0.80. In segmentation, YOLO overestimates a 13.5% of the predicted area on average, whereas Mask R-CNN overestimates a 9.2%. The proposed methods present a cost-effective solution for forest monitoring using RGB images and have been successfully used to monitor ∼146,500 acres of pine cultivations.