Tree detection and diameter estimation based on deep learning

Abstract

Tree perception is an essential building block toward autonomous forestry operations. Current developments generally consider input data from lidar sensors to solve forest navigation, tree detection and diameter estimation problems, whereas cameras paired with deep learning algorithms usually address species classification or forest anomaly detection. In either of these cases, data unavailability and forest diversity restrain deep learning developments for autonomous systems. Therefore, we propose two densely annotated image datasets—43 k synthetic, 100 real—for bounding box, segmentation mask and keypoint detections to assess the potential of vision-based methods. Deep neural network models trained on our datasets achieve a precision of 90.4 % for tree detection, 87.2 % for tree segmentation and centimeter accurate keypoint estimations. We measure our models’ generalizability when testing it on other forest datasets, and their scalability with different dataset sizes and architectural improvements. Overall, the experimental results offer promising avenues toward autonomous tree felling operations and other applied forestry problems. The datasets and pre-trained models in this article are publicly available on GitHub (https://github.com/norlab-ulaval/PercepTreeV1).

Low-Cost Non-Contact Forest Inventory: A Case Study of Kieni Forest in Kenya

“…Segmentation of the images in this study involved significant human interaction using image labelling software to ensure perfectly accurate masks were generated. This step can, however, be automated by implementing semantic segmentation based on deep learning [86,87], thus paving the way for real-time tree parameter estimation. This is an area of investigation for future research.…”

From Bark to Byte: Automating Forest Inventory Data Collection Through Camera and Mobile LIDAR

“…These datasets which have been previously collected and validated serve as a benchmark, which enables model training without manually producing tagged data where appropriate as well as comparison of results. Existing tagged tree-level RGB photo [3,10] and point cloud [33,35] data will be used as references for methodology, training, and evaluation where possible. By comparing the outcomes derived from the proposed approach with these established datasets, discrepancies can be identified and addressed.…”

An Advanced Software Platform and Algorithmic Framework for Mobile DBH Data Acquisition

“…Danilo et al [22] imporved the U-Net by adding a deep residual module for tree segmentation in urban environments and further fitting the skeleton model of trees. In addition, Vincent [32] used Mask R-CNN and Cascade Mask R-CNN for segmenting tree images and predicting keypoint locations, in which they reported tree detection rates and segmentation accuracies of 90.4% and 87.2%, respectively. Although both studies successfully extracted the trunk masks of trees from 2D images, their algorithms still have limitations in terms of real-time performance.…”

“…In order to accurately measure the tree's diameter at breast height (DBH), the acquired trunk region montages need to be further processed. The diameter-at-breast points, usually located at a height of about 1.3 m above the ground in the trunk, are two points in the tree cross-section that intersect the line connecting the center of mass [42]. The positioning accuracy of these points directly affects the accuracy of breast diameter measurements.…”

“…However, due to the drift characteristics of the SLAM algorithm, long time measurements may lead to cumulative errors. Grondin [32] optimized Mask R-CNN by adding a key point prediction module for predicting key points on the tree trunk, but this method requires a large amount of data annotation. They have a prediction error of 5.2 pixels for the diameter points of the thorax, and this error is further amplified during the mapping of the 2D image coordinate system to the 3D world coordinate system, resulting in a large deviation in the measured thorax values.…”