Urban Vegetation Mapping from Aerial Imagery Using Explainable AI (XAI)

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• Accurate urban vegetation mapping is vital for biodiversity, ecological balance, and mitigating urban heat islands.
• Traditional classification methods struggle with complex urban vegetation due to spatial and spectral similarities.
• Deep neural networks (DNNs) show promise but face challenges with feature relevance and labeled data quality.

Purpose of the Study:

• To improve urban vegetation mapping accuracy using Explainable AI (XAI).
• To interpret DNN model outputs for vegetation classification using Shapley additive explanations (SHAP).
• To identify and rank important spectral and textural features for vegetation mapping.

Main Methods:

• Applied a DNN model for vegetation cover mapping from aerial imagery.
• Utilized Shapley additive explanations (SHAP) to interpret the DNN model's decision-making process.
• Incorporated both spectral and textural features, including GLCM metrics, for enhanced classification.

Main Results:

• Achieved an overall accuracy of 94.44% for vegetation cover mapping.
• SHAP analysis identified Hue, Brightness, GLCM_Dissimilarity, GLCM_Homogeneity, and GLCM_Mean as highly influential features.
• Demonstrated the limitations of spectral-only approaches for vegetation classification.

Conclusions:

• Explainable AI (XAI) and SHAP effectively interpret DNN models for vegetation mapping.
• Texture features significantly enhance the accuracy of vegetation classification from aerial imagery.
• Integrating spectral and textural data with XAI provides a more robust approach to urban vegetation mapping.