DCA-UNet: A Cross-Modal Ginkgo Crown Recognition Method Based on Multi-Source Data.

Area of Science:

• Botany and Conservation Biology
• Computer Science and Artificial Intelligence
• Remote Sensing and Geospatial Analysis

Background:

• Wild ginkgo is an endangered species crucial for genetic resource conservation.
• Traditional surveys and satellite remote sensing have limitations in monitoring wild ginkgo in complex terrains.
• Existing deep learning models struggle with multimodal data fusion for accurate ginkgo crown recognition.

Purpose of the Study:

• To develop a precise method for wild ginkgo crown segmentation using UAV-based multimodal imagery.
• To propose a novel deep learning network that effectively fuses RGB and multispectral data.
• To enhance the recognition performance and generalization capability for monitoring endangered tree species.

Main Methods:

• A multimodal ginkgo crown dataset was created using UAV-acquired RGB and multispectral images.
• A dual-branch dynamic weighting fusion network, DCA-UNet, was proposed.
• DCA-UNet features a dual-branch encoder for independent feature extraction, a cross-modal interaction fusion module with attention, and an attention-enhanced decoder.

Main Results:

• DCA-UNet achieved high segmentation accuracy: 93.42% IoU, 96.82% PA, 96.38% Precision, and 96.60% F1-score.
• The proposed model significantly outperformed the DFAFNet and single-modality baseline models.
• The model demonstrated strong generalization and robustness across different flight altitudes and complex scenarios.

Conclusions:

• DCA-UNet offers a superior and efficient solution for UAV-based multimodal ginkgo crown recognition.
• The developed method provides a reliable tool for monitoring wild endangered tree species.
• Effective fusion of multimodal data significantly improves the accuracy and robustness of remote sensing-based species identification.