Machine learning approaches to forest species classification using spectral analysis.

Area of Science:

• Ecology
• Remote Sensing
• Forestry

Background:

• Ecosystem monitoring and species classification are enhanced by high-resolution remote sensing.
• Unmanned Aerial Systems (UAS) offer a cost-effective approach for ecological research.

Purpose of the Study:

• To evaluate Unmanned Aerial Systems (UAS) and machine learning for classifying dominant tree species in a Bottomland Hardwood Forest (BHF).
• To compare the effectiveness of a Convolutional Neural Network (CNN) and an Object-based Image Analysis (OBIA) approach for tree species classification.

Main Methods:

• Collected high-resolution RGB aerial imagery using UAS.
• Processed imagery with photogrammetric techniques to create orthomosaics and texture features.
• Applied a U-Net Convolutional Neural Network (CNN) and a Random Forest classifier via Object-based Image Analysis (OBIA) segmentation.

Main Results:

• The CNN approach achieved higher classification accuracy than OBIA for nine dominant tree species.
• Quercus species demonstrated the highest precision at 83.3% using the CNN method.
• The study identified potential for UAS and machine learning in forest species inventory and ecological monitoring.

Conclusions:

• UAS imagery combined with machine learning, particularly CNNs, shows significant promise for accurate tree species classification.
• Further research integrating multi-temporal data and advanced sensors can improve classification accuracy and practical applications in environmental monitoring.
• Findings support the increasing utility of UAS remote sensing in ecological research and forest conservation.