Remote sensing-based measurement of Living Environment Deprivation: Improving classical approaches with machine...

Area of Science:

• Urban studies
• Geospatial analysis
• Socioeconomic modeling

Background:

• Gathering socioeconomic data in urban settlements is crucial for policy and planning.
• Traditional methods can be time-consuming and costly.
• Very high spatial resolution (VHR) satellite imagery offers a potential alternative for data acquisition.

Purpose of the Study:

• To assess the utility of VHR imagery features for predicting socioeconomic indicators.
• To compare the performance of machine learning models against classical statistical approaches for socioeconomic index estimation.
• To introduce advanced machine learning techniques for enhanced predictive accuracy in remote sensing applications.

Main Methods:

• Extraction of land cover, spectral, structure, and texture features from VHR Google Earth imagery.
• Application of Ordinary Least Squares (OLS) regression and spatial lag models.
• Implementation and evaluation of Gradient Boost Regressor and Random Forests machine learning algorithms.
• Utilizing feature importance, partial dependence plots, and cross-validation for model interpretation and assessment.

Main Results:

• Random Forests achieved the highest predictive performance with an R-squared of approximately 0.54.
• Gradient Boost Regressor showed strong performance with an R-squared of 0.5.
• Classical methods, OLS (R2=0.3) and spatial lag model (R2=0.43), exhibited significantly lower predictive accuracy.

Conclusions:

• VHR imagery is a valuable data source for estimating socioeconomic conditions at a fine spatial scale.
• Machine learning models, especially Random Forests, significantly outperform traditional methods in predicting Living Environment Deprivation.
• The study demonstrates the potential of integrating advanced machine learning with remote sensing for socioeconomic analysis.