Utilizing active learning and attention-CNN to classify vegetation based on UAV multispectral data.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Ecology

Background:

• Accurate vegetation type identification is crucial for ecological monitoring.
• Traditional methods often require extensive labeled data, increasing costs.
• Unmanned Aerial Vehicle (UAV) remote sensing offers efficient data acquisition.

Purpose of the Study:

• To develop a deep learning model for accurate vegetation classification using UAV multispectral data.
• To implement an active learning strategy to minimize data labeling costs.
• To enhance the model's feature extraction capabilities for distinguishing similar vegetation types.

Main Methods:

• Utilized multispectral data from UAV remote sensing.
• Employed an active learning strategy with minimum confidence scoring and data pool sampling.
• Developed a deep learning model featuring a semantic segmentation gated full fusion module with a dual attention mechanism.

Main Results:

• Achieved 93.2% average accuracy at 20% labeling cost.
• Demonstrated superior classification accuracy compared to other models with limited training samples.
• Reached 95.32% average accuracy at full annotation cost, with only a 2% difference while saving 80% on annotation costs.

Conclusions:

• Active learning effectively identifies high-value samples, significantly reducing annotation costs.
• The proposed model enhances vegetation classification accuracy and efficiency.
• Field investigations validated the model's reliable performance in identifying surface vegetation cover types.