Methodological Benchmark for Machine Learning in Urban Expansion Monitoring in Arid Regions: Taking Riyadh as an Example

Crossing Spectral Haze: Machine Learning-Driven Monitoring and Methodological Reconstruction of Arid Urban Expansion

In the 21st century, urbanization is a decisive force reshaping the global landscape. Especially in arid and semi-arid regions, this expansion faces unique engineering challenges: sparse vegetation and high similarity in surface reflectance create significant "spectral mimicry" between built-up areas (such as concrete and asphalt) and bare desert soils in the spectral domain. This ambiguity not only poses a systemic threat to traditional Land Use/Land Cover (LULC) classification methods but also puts stringent tests on long-term national planning and resource allocation.

Saudi Arabia, particularly the capital Riyadh, is at the forefront of this grand transformation. Driven by "Saudi Vision 2030," Riyadh is undergoing explosive urbanization. According to research data, the growth in built-up area from 416 km² in 1990 to 1,219 km² in 2025 is astonishing, reflecting not only population growth but also the country's strategy for economic diversification and concentrated infrastructure investment. However, this rapid, non-linear spatial expansion places higher demands on traditional monitoring and planning systems.

Challenge: Spectral Feature Separation in Arid Environments

When monitoring urban changes using remote sensing, effectively distinguishing urban structures from natural topography is the core bottleneck. Satellite imagery in arid environments like Riyadh has spectral signatures that make it difficult to differentiate high-reflectance building materials from low-reflectance bare soils. Previous analyses often rely on single indices (such as NDBI), but such a single indicator often lacks sufficient discriminatory power under complex multi-scale, multi-temporal changes, making the robustness of the classification results difficult to guarantee.

To address this systemic methodological flaw, this study introduces non-parametric classifiers from Machine Learning (ML) into the remote sensing monitoring field. The innovation of this research lies in moving beyond reliance on single, empirical indices to construct a Custom Feature Stack that integrates key indicators such as the Bare Soil Index (BSI) and NDBI, aiming to provide finer, separable feature dimensions for complex arid landscapes.

Engineering Capital Perspective: Systematic Benchmark Testing of Model Selection

For infrastructure investment and long-term planning, choosing the right model is key to determining project risk and monitoring accuracy.### Engineering Capital Perspective: Systematic Benchmarking of Model Selection

For infrastructure investment and long-term planning, choosing the right model is key to determining project risk and monitoring accuracy. This study goes beyond simple model application; its core value lies in establishing a cross-algorithm benchmarking framework. The study systematically compared five mainstream supervised learning algorithms—Random Forest (RF), Support Vector Machine (SVM), Gradient Boosted Trees (GBT), Classification and Regression Trees (CART), and k-Nearest Neighbor (KNN)—under rigorous testing.

The significance of this systematic comparison is: it forces analysts to move beyond "trial and error" and identify which algorithm architecture (such as the non-linear fitting capability of RF) is most effective at capturing the non-linear trends of urban expansion under specific drought spectral noise. The research results clearly indicate that the Random Forest model performed best in terms of overall accuracy (Overall Accuracy = 0.977) and Kappa coefficient (Kappa = 0.954), providing a highly reliable engineering practice path for future Land Use/Land Cover (LULC) classification in similar arid environments.

Mapping Regional Development and Long-Term Trends

From the perspective of engineering capital flow, high-precision, reproducible urban expansion data is a prerequisite for attracting long-term infrastructure investment (whether public or PPP projects). When national strategies (such as Saudi Vision 2030) have clear quantitative targets for urban morphology, precise spatial dynamic assessments based on long-term time series (1990-2025) can effectively reduce investment risk and guide resource allocation towards high-growth areas.

This study not only provides a high-precision spatial dataset but, more importantly, it offers a transferable methodological framework. This framework is applicable not only to Riyadh but can be generalized to the urban monitoring challenges in arid and semi-arid regions globally, providing important engineering and policy references for the Global South on how to leverage data science to enhance the scientific rigor and foresight of infrastructure planning amidst rapid urbanization.

Conclusion: Monitoring urban expansion has evolved from simple image recognition to a complex interdisciplinary engineering problem. By combining advanced machine learning techniques with rigorous remote sensing methodologies, we have not only quantified Riyadh's urbanization achievements but also constructed an engineering analytical paradigm to address the urban governance challenges in arid regions worldwide.