Vegetation extraction through UAV RGB imagery and efficient feature selection.

Area of Science:

• Ecological remote sensing
• Environmental monitoring
• Geospatial analysis

Background:

• Accurate vegetation identification is critical for ecological assessments and monitoring in mining areas.
• Highly heterogeneous vegetation in mining sites presents significant challenges for detailed mapping.
• Feature selection and combination are key to improving vegetation classification accuracy.

Purpose of the Study:

• To develop an effective method for fine vegetation mapping in heterogeneous mining areas.
• To evaluate the importance of different features (vegetation indexes and textures) for vegetation identification.
• To compare a novel feature selection method with traditional approaches.

Main Methods:

• Utilized Unmanned Aerial Vehicle (UAV) RGB imagery to extract vegetation indexes and texture features.
• Selected optimal features using standard deviation and difference coefficient criteria.
• Classified vegetation using Support Vector Machine (SVM) with various feature combinations.

Main Results:

• The combination of RGB images with all selected features achieved the highest accuracy (OA up to 88.74%, Kappa up to 0.8505).
• Texture features were found to be more influential than vegetation indexes for vegetation identification in this context.
• The difference coefficient-based feature selection method proved simpler and more effective than Random Forest (RF).

Conclusions:

• The proposed method, integrating selected features with UAV RGB imagery and SVM classification, is effective for mapping heterogeneous vegetation in mining areas.
• Texture features play a crucial role in distinguishing vegetation types in complex environments.
• The difference coefficient method offers a simple yet powerful approach for feature selection in ecological mapping.