Slum Detection and Density Mapping with AlphaEarth Foundations: A Representation Learning Evaluation Across 12 Global Cities

Abstract: Pixel-level slum mapping has long been constrained by limited cross-city generalisation, the absence of continuous density estimation, and weak global comparability. AlphaEarth Foundations (AEF), a globally consistent 64-dimensional annual surface embedding at 10 m, offers a new analysis-ready basis for lightweight slum monitoring, but its applicability to slum detection - an indirectly coupled task shaped by both built form and socio-economic processes - remains untested. We evaluate AEF on slum classification and sub-pixel density estimation across 12 cities and 69 city-year pairs (2017-2024), using GRAM pseudo-masks as supervisory labels. The evaluation spans four training strategies, two protocols (random split and 3x3 spatial block cross-validation), six auxiliary feature configurations, and five baseline models, complemented by representation-level analyses (PCA, SHAP) and full-AOI mapping. Five findings emerge. (1) Same-city cross-year training is optimal under both protocols (median spatial F1 = 0.616, R^2 = 0.466); temporal expansion outperforms cross-city transfer, indicating city-scale representational drift. (2) Regression R^2 is driven primarily by zero/non-zero boundary discrimination: positive-pixel R^2 is consistently negative across all cities, revealing limited capacity to model intra-pixel density gradients at 10 m. (3) PC36 is consistently top-ranked across tasks; classification saturates at k = 32 while regression remains unsaturated at k = 64. (4) POI features yield the largest density gain (Delta R^2 = +0.064). (5) For six cities meeting dual-task usability thresholds, full-AOI inference across 2017-2024 preserves slum cluster structure (mean SSIM = 0.926). The study delineates the capabilities and complementarity needs of foundation-model embeddings for slum monitoring.

• The paper demonstrates that same-city cross-year training achieves optimal spatial F1 (0.616) while revealing severe representation drift during cross-city transfer.
• The study rigorously compares five modeling approaches and auxiliary features, highlighting the dominant role of PC36 and the limited benefit of nighttime lights.
• The research underscores scalable slum boundary mapping while identifying major challenges in fine-grained density regression and spatial error clustering.

Systematic Evaluation of AlphaEarth Foundations Embeddings for Slum Detection and Sub-Pixel Density Mapping

Introduction

The paper "Slum Detection and Density Mapping with AlphaEarth Foundations: A Representation Learning Evaluation Across 12 Global Cities" (2605.10029) addresses the challenge of scalable, globally consistent slum mapping by leveraging AlphaEarth Foundations (AEF), a 64-dimensional, 10 m annual surface embedding product. The study presents an extensive evaluation across dual tasks—slum boundary classification and sub-pixel density regression—using data from 12 cities. The research systematically quantifies the representational suitability of AEF embeddings, investigates auxiliary feature contributions, characterizes dimensional structure, and benchmarks spatial generalization and application robustness, thus providing the most comprehensive foundation-model-based slum mapping evaluation to date.

Methodology

Study Design

The evaluation corpus spans 69 city-year pairs across Africa, Latin America, and Asia, incorporating major urban centers that collectively represent diverse slum morphologies and varying positive class proportions (from 1.4% to 22.6%). Supervisory labels are derived from the GRAM pseudo-mask slum segmentation, downsampled from ~0.6 m to a 10 m pixel grid.

The paper benchmarks five model families (logistic regression, ridge regression, random forest, gradient-boosted trees, 4-layer MLP), using four training schemes varying spatial and temporal transfer: single-city/year, single-city/cross-year, cross-city/same year, and cross-city/cross-year. Auxiliary feature stacks—nighttime lights (NTL), remote sensing (RS), spatial proximity, and points-of-interest (POI)—are concatenated to the AEF feature vectors in various combinations to assess their marginal utility.

Dual-task evaluation (classification and regression) utilizes both random splits and strict spatial block cross-validation to rigorously assess generalization, with additional representation structure analysis via PCA and SHAP.

Main Results

Benchmark Model and Transfer Protocol Performance

The study demonstrates that cross-year training within the same city (S2 protocol) is the optimal configuration, yielding a median spatial F1 of 0.616 (classification) and $R^2$ of 0.466 (regression) for TorchMLP. Cross-city transfer leads to significant performance degradation, indicating strong city-scale representation drift within AlphaEarth embeddings.

TorchMLP consistently outperforms tree-based and linear models in both tasks, but all models suffer pronounced accuracy loss under spatial block validation compared to random splits, emphasizing the need for spatial independence in remote sensing generalization studies.

Regression Task Decomposition

A central empirical finding is that regression performance is primarily driven by accurate discrimination between slum-present and slum-absent pixels. For positive pixels (i.e., those containing slum at sub-pixel level), all models yield consistently negative $R^2$ scores (median −2.00), underscoring AEF’s limitations for fine-grained intra-pixel density estimation at 10 m.

Representation Structure

Dimensional ablation (PCA) and SHAP analysis reveal that PC36 is the dominant principal component across both tasks; classification saturates at 32 principal components, while regression remains unsaturated even at 64 components, indicating that density mapping is more dependent on broader, possibly less information-dense, embedding subspaces. Importantly, secondary dimensions are non-overlapping between classification and regression, supporting differentiated embedding utility for distinct tasks.

Auxiliary Features

Among auxiliary stacks, POI yields the largest marginal gain for density regression ($\Delta R^2 = +0.064$), and a combined feature stack (AEF+NTL+RS+Spatial+POI, denoted C5) offers the best aggregate classification performance but at the cost of increased inter-city heterogeneity and some negative interactions. Notably, NTL offers negligible (and occasionally negative) discriminative value, which is attributed to the low-light nature of slum settlements and confounding with other urban deprivation contexts.

Application Robustness and Spatial Structural Fidelity

For the subset of cities meeting dual-task usability thresholds (F1 ≥ 0.5, $R^2$ > 0.3; six cities in total), full-scene inference from 2017 to 2024 produces topological maps with high structural fidelity—mean SSIM = 0.926—and preserves core High-High slum clusters in LISA analysis. However, residual spatial error is non-negligible and clustered (residual Moran’s I > 0.2 in all cities), indicating that prediction failures remain spatially structured and require context-aware correction.

Implications and Theoretical Insights

Practical Utility

The results validate the utility of AEF for slum boundary detection and coarse monitoring applications such as UN SDG 11.1 targeting, offering a scalable alternative to computationally intensive very-high-resolution (VHR) solutions. In high-prevalence, morphologically archetypal cities, the framework delivers reliable spatial and temporal inference for urban planning, resource allocation, and longitudinal monitoring.

The inability to recover intra-pixel density gradients at 10 m, however, places a hard limit on the suitability of AlphaEarth embeddings for fine-grained service radius estimation, detailed population or infrastructure modeling, and planning in cities with highly atypical slum structures or extreme class imbalance.

Model and Representation Generalization

The pronounced city-scale representation drift in AEF restricts the viability of global or cross-city model transfer and underscores the necessity for city-aware or adaptive downstream architectures. The findings suggest foundational geospatial models, when applied to indirectly coupled and highly heterogeneous tasks such as slum mapping, require careful transfer calibration and possibly explicit city- or region-specific adaptation modules.

Feature Integration

The varied marginal utility and occasional negative interaction observed among auxiliary feature types suggest that naïve concatenation is suboptimal for harnessing complementary semantic signals, especially where city heterogeneity or slum morphological diversity is high. There is a clear need for structured cross-modal fusion mechanisms (e.g., gated attention, alignment networks) to maximize downstream utility.

Limitations and Future Directions

Several limitations arise: (1) All evaluation utilizes GRAM pseudo-masks as reference, not independent field data; (2) Auxiliary feature fusion is naïve and lacks adaptive interplay modeling; (3) Five cities fall below usability thresholds, indicating that representation limits and city-level morphology constrain applicability; (4) Residual predictions exhibit spatial clustering, which is unmitigated by the current modeling paradigm.

Future technical priorities include: advanced feature fusion schemes; mixture-of-experts or city-adaptive downstream routing; targeted integration of higher resolution or auxiliary imagery for fine gradient tasks; and explicit spatial regularization or graph-based modeling to address correlated error.

Conclusion

This study delivers the first rigorous, cross-city evaluation of foundation-model-based representation learning for pixel-scale slum classification and density mapping. While AEF embeddings at 10 m enable reproducible and scalable slum boundary detection in global applications, significant constraints persist for density regression and cross-city transfer, with practical accuracy tightly coupled to the interaction of slum prevalence, morphology, and auxiliary feature availability. These empirical results advance the quantitative understanding of foundation geospatial model limitations and set new benchmarks for actionable slum monitoring using public embeddings and lightweight models.

Future advances in adaptive model architectures, multi-source fusion, and resolution-sensitive representation learning will further enhance the practical impact of foundation models in urban poverty mapping and globally standardized monitoring.