Beyond AlphaEarth: Toward Human-Centered Geospatial Foundation Models via POI-Guided Contrastive Learning

1. Problem Statement

Geospatial Foundation Models (GFMs), such as AlphaEarth (AE), have successfully generated high-resolution (10m) embeddings from multi-source Earth Observation (EO) data. These models excel at capturing physical and spectral patterns (e.g., land cover, vegetation, topography). However, they face significant limitations in urban contexts:

• Lack of Functional Semantics: AE embeddings primarily encode physical signals, failing to capture human activities, socioeconomic structures, or functional urban organization (e.g., distinguishing a "shopping district" from a "residential area" based solely on spectral data).
• Poor Interpretability: The latent representations are difficult to query or interpret using natural language, limiting their utility for open-vocabulary spatial analysis.
• Data Gaps: Existing Points of Interest (POI) models capture human semantics but lack explicit spatial structure and high-resolution continuity.

The core challenge is how to align physically grounded EO representations with human-centered semantic signals to create representations that are both functionally expressive and language-accessible.

2. Methodology: AETHER Framework

The authors propose AETHER (AlphaEarth–POI Enriched Representation Learning), a lightweight multimodal contrastive alignment framework. The goal is to adapt AlphaEarth embeddings to urban semantics using POIs as a bridge.

Key Components:

1. POI Text Encoder:
   • POIs are enriched with hierarchical descriptions (Name + Category + Subcategory) using a template (e.g., "A place of [subcat], a type of [cat], named [name]").
   • A pretrained Large Language Model (LLM) encodes these texts into embeddings, which are then projected into a shared latent space.
2. AlphaEarth (AE) Encoder:
   • For each POI location, the framework aggregates AE embeddings from the surrounding area using multi-scale spatial windows (a base view and an augmented view).
   • This aggregation mitigates the mismatch between point-based POI annotations and the spatial extent of real-world urban objects.
   • A lightweight projector maps these aggregated AE features into the shared latent space.
3. Contrastive Alignment Module:
   The model is trained using a joint objective function combining two symmetric contrastive losses:
   • Cross-Modal Alignment ($L_{AP}$): Aligns the AE embedding of a location with the text embedding of the POI at that location. This grounds the physical features in human semantics.
   • Intra-Modal Consistency ($L_{AA}$): Enforces consistency between the base-view and augmented-view AE embeddings of the same location. This acts as a regularizer to ensure spatial stability across scales.
   • Total Loss: $L = \lambda L_{AA} + (1-\lambda) L_{AP}$, where $\lambda$ balances the two objectives.

Inference:

During inference, the text encoder is discarded. The trained AE projector is applied to the entire city's AE embedding field to generate semantically enriched spatial embeddings. These embeddings can be used for:

• Supervised downstream tasks (classification, regression).
• Natural Language-Conditioned Spatial Retrieval: Users can query the map with text (e.g., "Green Park") to retrieve spatially relevant regions based on cosine similarity.

3. Key Contributions

• Bridging EO and Urban Semantics: AETHER is the first framework to connect physically grounded AlphaEarth embeddings with human-centered semantic signals via POI-guided contrastive alignment.
• POI-Guided Multimodal Framework: It introduces a lightweight architecture that aggregates multi-scale spatial features and aligns them with text embeddings, effectively integrating functional semantics without losing spatial continuity.
• Enhanced Interpretability: The resulting embedding space is "language-accessible," enabling open-vocabulary spatial retrieval without task-specific supervision.
• Comprehensive Validation: The framework is validated across two distinct cities (London and Singapore) and four diverse tasks, demonstrating robustness and state-of-the-art performance.

4. Experimental Results

The study was evaluated in Greater London and Singapore across four tasks:

1. Land-Use Classification (LUC): Discrete classification of land use.
2. Socioeconomic Distribution Mapping (SDM): Predicting occupational distributions.
3. Land-Use Distribution Mapping (LUD): Predicting proportional land-use compositions.
4. GDP Prediction: Scalar regression of GDP per capita.

Performance Highlights:

• State-of-the-Art (SOTA): AETHER outperformed all baselines (including AlphaEarth, CaLLiPer, SatCLIP, and random baselines) across all tasks.
• Improvement Magnitude: Relative improvements ranged from 4.5% to 21.9%.
  • London LUC: +6.3% macro-F1 over AlphaEarth.
  • Singapore GDP: +12.7% reduction in MAE over the best baseline.
  • SDM/LUD: Significant reductions in KL divergence and distributional errors, particularly for socially grounded tasks where EO data alone is insufficient.
• Spatial Retrieval:
  • Qualitative: The model successfully localized functional regions (e.g., CBDs, airports, parks) using natural language queries, capturing both semantic meaning and spatial boundaries.
  • Quantitative: In retrieval tasks (Recall@Top-k%), AETHER significantly outperformed CaLLiPer and AlphaEarth (e.g., 54.56% Recall@Top-1% for "Art" vs. 16.39% for CaLLiPer).
• Efficiency: The model is highly efficient, with only 0.17M trainable parameters and a training time of ~4.7 seconds per epoch on a high-end GPU.

5. Significance and Implications

• From Physical to Functional: The paper demonstrates that GFMs can be evolved from purely physical descriptors to human-centered representations by injecting functional semantics.
• Semantic Propagation: A key finding is that semantic alignment does not just improve performance at POI locations; the multi-scale design allows functional signals to propagate to surrounding areas, improving predictions even in regions with sparse POI data.
• Language-Accessible GIS: AETHER paves the way for interactive geospatial foundation models. It enables users to query complex spatial patterns using natural language, moving beyond rigid, predefined analytical pipelines to flexible, open-vocabulary exploration.
• Complementarity: The results confirm that EO data provides spatial completeness and morphological context, while POI data provides the functional richness necessary for understanding human systems. Their integration creates a more holistic view of the urban environment.

In conclusion, AETHER represents a significant step toward making geospatial foundation models interpretable, functionally aware, and accessible to human users through natural language.