Deep learning-based dead pine tree detection from unmanned aerial vehicle images

Understanding the spatial distribution of dead pine trees (DPTs) in mountainous areas is vital for diseased wood cleanup and the prediction of pine wilt disease. The study induced a deep learning (DL) model of convolutional neural networks (CNNs) for the DPT detection using unmanned aerial vehicle i...

Full description

Saved in:
Bibliographic Details
Published in:International journal of remote sensing 2020-11, Vol.41 (21), p.8238-8255
Main Authors: Tao, Huan, Li, Cunjun, Zhao, Dan, Deng, Shiqing, Hu, Haitang, Xu, Xinluo, Jing, Weibin
Format: Article
Language:English
Subjects:
Accuracy
Artificial neural networks
Classification
Deep learning
Detection
Environmental cleanup
Iterative methods
Machine learning
Model accuracy
Mountainous areas
Neural networks
Pine
Pine trees
Spatial distribution
Template matching
Training
Trees
Unmanned aerial vehicles
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Understanding the spatial distribution of dead pine trees (DPTs) in mountainous areas is vital for diseased wood cleanup and the prediction of pine wilt disease. The study induced a deep learning (DL) model of convolutional neural networks (CNNs) for the DPT detection using unmanned aerial vehicle images. Two thousand manually labelled DPT and Non-DPT samples were collected to train (80%) and optimize (20%) the CNNs of AlexNet and GoogLeNet, and 768 samples from other three areas were used to predict the categories in Jinjiang, Fujian, southeastern China. The same dataset was used to compare the classification accuracy between CNNs and the traditional template matching (TM) method. In addition, the potential factors influencing the overall accuracy of the DL model for DPT detection were evaluated. Training results of AlexNet and GoogLeNet showed an accuracy of over 93.30% and 97.38% respectively for the top-1 test after the final training iteration. For the DPT classification in the prediction areas of pure forests, the TM method showed the 56-65% of overall accuracy while CNNs showed 65-80%. CNNs-based DPT detection was more accurate than the traditional TM method. Terrains and red broadleaf trees influence the CNNs-based DPT detection from aerial visible-images in mixed artificial forest and peaks areas.
ISSN:0143-1161
1366-5901
DOI:10.1080/01431161.2020.1766145