Explainable few-shot learning workflow for detecting invasive and exotic tree species.

Area of Science:

• Ecology
• Computer Science
• Remote Sensing

Background:

• Deep Learning models require large labeled datasets, posing challenges for new applications with limited data.
• Few-shot learning addresses data scarcity but often lacks sufficient explanation for model predictions.
• Accurate identification of invasive tree species is crucial for forest management and biodiversity conservation.

Purpose of the Study:

• To propose an explainable few-shot learning workflow for detecting invasive and exotic tree species.
• To integrate Siamese networks with explainable AI (XAI) for robust tree species classification.
• To provide visual, case-based explanations for model predictions in data-scarce environments.

Main Methods:

• Developed a workflow combining Siamese networks with XAI techniques.
• Utilized Unmanned Aerial Vehicle (UAV) imagery for tree species detection.
• Employed a lightweight backbone (MobileNet) for efficient model training.

Main Results:

• Achieved an F1-score of 0.86 in 3-shot learning, outperforming a shallow Convolutional Neural Network (CNN).
• Demonstrated effective classification of tree species even under data-scarce conditions.
• Provided explanation metrics (correctness, continuity, contrastivity) and visual cases for prediction insights.

Conclusions:

• The proposed workflow successfully addresses challenges of data scarcity and lack of explainability in tree species detection.
• This approach enhances the application of AI and UAVs in forest management, biodiversity conservation, and the study of rare species.