The role of ExtSpecR in streamlining UAV-based tree phenomics and spectral analysis

[EmperoR]

 

Due to their ability to collect tree phenotypic trait data in large quantities, unmanned aerial vehicles, or UAVs, have completely changed the forestry industry. Even with the progress made in object detection and remote sensing, precise identification and extraction of spectral data for individual trees continue to be major obstacles, frequently necessitating tedious manual annotation.

For better tree detection, current research focuses on developing segmentation algorithms and convolutional neural networks; however, the requirement for precise manual labeling prevents these technologies from being widely adopted. This emphasizes how critical it is to create a higher-throughput, more effective technique for automatically extracting spectral information for individual trees. 
The open-source tool ExtSpecR, which offers an intuitive interactive web application, is presented in this paper as a means of achieving single tree spectral extraction in forestry using UAV-based imagery. It optimizes the process of spectral and spatial feature extraction by speeding up the identification and annotation of individual trees. 
Users can calculate vegetation indices and view outputs as false-color and VI-specific images by uploading TIFF-formatted spectral images through the ExtSpecR user interface. Users upload point cloud data and multispectral images to the interactive dashboard, which then defines the region of interest (ROI) for tree identification and segmentation, enabling the system's core phenotyping capabilities. 
This procedure produces 3D visualizations of the segmented trees by utilizing the lidR package's "locate_trees" function. Evaluation of ExtSpecR's performance in comparison to ground truth in tree plantations with different canopy densities shows that it can detect individual trees with accuracy ranging from 91% to 97%. 
By comparing ExtSpecR's functionality to that of other tools, its distinct approach of fusing point cloud data and multispectral imagery with already-existing algorithms for optimal user experience and thorough tree analysis is highlighted. 
For better outcomes, recommendations include segmenting point cloud data and defining specific target areas, even though it faces difficulties with large input data sizes and complex environments with overlapping canopies. Further improvements, according to the paper, ought to focus on raising cloud quality and assessing effectiveness using hyperspectral imagery and LiDAR point clouds. 

page:

GitHub - Yanjie-Li/ExtSpecR: Tree detection, segementation and spectral extraction