Conference Agenda

General-purpose spatial representations are essential for building transferable geospatial foundation models (GFMs). Among them, the AlphaEarth Foundation (AE) represents a major step toward a global, unified representation of the Earth's surface, learning 10-meter embeddings from multi-source Earth Observation (EO) data that capture rich physical and environmental patterns across diverse landscapes. However, such EO-driven representations remain limited in capturing the functional and socioeconomic dimensions of cities, as they primarily encode physical and spectral patterns rather than human activities or spatial functions. We propose AETHER(AlphaEarth–POI Enriched Representation Learning), a lightweight framework that adapts AlphaEarth to human-centered urban analysis through multimodal alignment guided by Points of Interest (POIs). AETHER aligns AE embeddings with textual representations of POIs, enriching physically grounded EO features with semantic cues about urban functions and socioeconomic contexts. In Greater London, AETHER achieves consistent gains over the AE baseline, with a 7.2% relative improvement in land-use classification F1 and a 23.6% relative reduction in Kullback–Leibler divergence for socioeconomic mapping. Built upon pretrained AE, AETHER leverages a lightweight multimodal alignment to enrich it with human-centered semantics while remaining computationally efficient and scalable for urban applications. By coupling EO with human-centered semantics, it advances geospatial foundation models toward general-purpose urban representations that integrate both physical form and functional meaning.

This work demonstrates the potential of foundation-model-encoded satellite embeddings to bridge Earth's surface and atmosphere. Based on our multimodal foundation model Copernicus-FM, we curate a global embedding dataset at 0.25°x0.25° resolutions (in consistency with ERA5). For each grid, multi-sensor images from Sentinel-1, 2, 3, 5P, and DEM are encoded into an embedding vector. These grid embeddings serve as condensed surface representations for downstream users. We verify their benefits as input for a climate task that predicts the 10-year mean and standard deviation of several climate parameters (e.g., precipitation) from ERA5. Compared to raw coordinates or location encodings, our results suggest that introducing surface embeddings helps produce more accurate prediction maps, reducing RMSEs by an average of up to 45%.

Ground-level horizon imagery is scarce across planetary bodies, making representation-centred approaches attractive for downstream geospatial tasks. We present DESPINA, a geospatial representation system that converts digital elevation models (DEMs) into structured neural embeddings of terrain geometry that condition a diffusion model to produce geometry-preserving, terrain-consistent visual reconstructions for a specified location and view direction.

Our pipeline integrates numeric elevation data (DEMs), structural embeddings (inverse-depth and soft edges), and textual priors, unifying heterogeneous geospatial signals into a shared, metric conditioning space. Using a Stable Diffusion model constrained with ControlNet, we can generate geologically consistent yet texturally diverse horizon datasets. Appearance priors are learned from historical surface photography to capture realistic textures and lighting cues, and geometric validation is performed against DEM-derived skylines and depth structure, independent of photographic training data. Through quantitative evaluation and a pilot qualitative study, DESPINA maintains skyline fidelity and geological boundaries while improving structural similarity relative to an image-conditioned baseline. Although our experiments use lunar DEMs and historical surface photography, the method is domain-agnostic and applicable to Earth, Mars, and other planetary DEMs.

Satellite -based crop classification and maps are important tools for food security and climate change mitigation, but existing approaches are not effective for small field systems. To address this, crop type classification using embeddings generated by a global foundation model, TESSERA, are compared to standard classification approaches in the literature. We find that our embedding -based approach offers a triple win: 1) consistent and statistically significant performance improvement over current methods, 2) greater simplicity due to the elimination of feature engineering, and 3) the reduction of computational cost. Our embedding -based approach achieves significantly higher F1 scores in the classification of 5 of 7 crop types for small fields in Austria (over 10% improvement in one case). Additionally, the TESSERA embedding -based method uses 8% of compute compared to the raw data method. These results indicate that embeddings are an effective approach for crop type classification tasks in small field systems.

This study explores the potential of the AlphaEarth Foundation (AEF) embeddings, a global, annual, analysis-ready satellite embedding dataset, for winter wheat and crop rotation mapping. Firstly, we analyze AEF embeddings for intra-class consistency and inter-class separability, assessing their effectiveness in representing wheat within the semantic embedding space. Subsequently, we compare multiple lightweight classifiers to identify an optimal model and conduct spatiotemporal generalization experiments across Henan Province from 2018 to 2024 using only a limited set of labelled samples from 2020. Based on the resulting wheat distribution maps, crop rotation patterns are further identified.Experimental results demonstrate that AEF embeddings exhibit strong semantic coherence and discriminative capability. Acceptable classification accuracy (OA = 0.85) can already be achieved using simple models such as cosine similarity and linear regression. More advanced lightweight classifiers further improve the performance (OA = 0.86–0.93) while maintaining stable results across different years and regions (spatial consistency = 0.82). In addition, the crop rotation maps show high spatial agreement with existing products, while producing more spatially contiguous field patterns.Overall, this study confirms that AEF embeddings can serve as effective, ready-to-use features for large-scale agricultural remote sensing applications. By substantially reducing the reliance on complex feature engineering and extensive training samples, they provide a practical and scalable solution for mapping winter wheat and its crop rotation patterns.