Individual Tree Species Classification Based on Convolutional Neural Networks and Multitemporal High-Resolution...

Area of Science:

• Forestry
• Remote Sensing
• Computer Science

Background:

• Individual tree species (ITS) classification is vital for forest management and protection.
• Airborne LiDAR and aerial photography offer high accuracy but are costly for large-scale applications.
• High-resolution satellite remote sensing data presents a viable alternative for extensive ITS classification.

Purpose of the Study:

• To assess the potential of high-resolution satellite imagery for ITS classification.
• To improve ITS classification accuracy by integrating multi-seasonal data using convolutional neural networks (CNNs).
• To compare the performance of different CNN models (DenseNet, ResNet, GoogLeNet) for this task.

Main Methods:

• Utilized WorldView-3 and Google Earth satellite imagery.
• Employed convolutional neural network (CNN) models, specifically DenseNet, ResNet, and GoogLeNet.
• Integrated feature information from different seasonal images to enhance classification.

Main Results:

• DenseNet outperformed ResNet and GoogLeNet in ITS classification.
• Achieved an Overall Accuracy (OA) of 75.1% for seven tree species using only WorldView-3 data.
• Improved OA to 78.1% by combining WorldView-3 and autumn Google Earth images.

Conclusions:

• CNN models, particularly DenseNet, are effective for ITS classification using satellite data.
• Integrating multi-seasonal Google Earth imagery as auxiliary data significantly enhances classification accuracy.
• Satellite remote sensing provides a scalable and cost-effective approach for large-area ITS classification.