Real estate valuation with multi-source image fusion and enhanced machine learning pipeline

Area of Science:

• Real Estate Valuation
• Machine Learning
• Computer Vision

Background:

• Automated valuation models (AVMs) commonly use structured data, often neglecting valuable unstructured information like images.
• Traditional property valuation methods can be subjective and lack accuracy.
• Previous machine learning (ML) approaches for real estate valuation have not fully explored multi-source image data integration or feature configuration optimization.

Purpose of the Study:

• To propose an enhanced ML-based real estate valuation framework integrating multi-source image data and feature configuration.
• To investigate the impact of different feature configurations on model performance for property valuation.
• To analyze the influence of fused image features on housing price prediction accuracy.

Main Methods:

• Utilized Hong Kong property data for a case study.
• Implemented an enhanced ML framework incorporating feature configuration and fusion of exterior, street view, and remote sensing images.
• Trained and compared eight ML regressors: Random Forest, Extra Tree, XGBoost, LightGBM, KNN, SVR, MLP, and MLR.
• Employed the SHapley Additive exPlanations (SHAP) method to interpret image feature impacts.

Main Results:

• Model performance varied significantly with different feature configurations, highlighting the importance of optimization.
• The Extra Tree regressor demonstrated superior performance compared to other tested models.
• Image features constituted half of the top 10 significant predictors, confirming their substantial contribution to valuation accuracy.
• Nonlinear relationships were observed between image features and housing prices, with distinct spatial patterns.

Conclusions:

• Feature configuration is crucial for developing accurate and reliable automated valuation models.
• Integrating multi-source image data significantly enhances property valuation accuracy.
• Understanding the nonlinear associations and spatial variations of image feature impacts is vital for urban planning and real estate development.