Deep learning-based palm tree detection in unmanned aerial vehicle imagery with Mask R-CNN

Abstract

Oil palm is highly valuable in tropical regions like Southeast Asia, including Indonesia. Therefore, accurate monitoring of oil palm trees is necessary for operational efficiency and reducing its environmental impact. Geospatial data, such as orthomosaic imagery from the unmanned aerial vehicle (UAV), can facilitate this goal. This research aims to integrate UAV data with deep learning algorithms, specifically Mask region-based convolutional neural network (R-CNN), to detect oil palm trees in Indonesia. We utilized Resnet-50 as the backbone and trained the model using data sampled from the template matching tool in eCognition. Considering factors like cloud shadows and other features, such as other plants, buildings, and road segments, we divided the study area into three containing different feature combinations in each. The Mask R-CNN model achieved an accuracy exceeding 80%, which is sufficient and makes it suitable for large-scale oil palm tree detection using high resolution images from UAV.