LoCA: Enabling Multi-Modality in Geospatial Location Encoders
Authors: Panagiotis Koromilas, Eleanna Vali, Diogenis Antonopoulos, Eleni Saka, Dimitris Christofis, Nikolaos Ioannis Bountos, Konstantin Klemmer, Ioannis Papoutsis, Yannis Panagakis, Mihalis A. Nicolaou
Organizations: University of Athens · The Cyprus Institute · Archimedes, Athena Research Center · National Technical University of Athens · Harokopio University · European Space Agency · University College London · LGND AI, Inc. · National Observatory of Athens · University of Cyprus
Location encoders map geographic coordinates to high-dimensional descriptors through contrastive alignment with paired Earth Observation imagery. We prove that this construction is structurally limited: any Lipschitz location encoder produces an output of intrinsic dimension at most 2, regardless of architecture or bandwidth, which forces the contrastive loss to strictly exceed its asymptotic optimum. Empirically, joint training also collapses the encoder's embedding covariance to low rank. Although dimension-limited, current location encoders sidestep this spectral collapse by training against a frozen image encoder, which we reinterpret as a covariance regularizer that supplies a high-rank target for the location encoder to fold onto. However, extending this approach to multimodal inputs is challenging, as location encoders should simultaneously map a single location to multiple, potentially incompatible, descriptors. We identify two requirements a regularizer must satisfy for location encoding in multimodal regimes: co-alignment across modalities, since adapters alone cannot align independently pretrained encoders; and high effective rank, since the location encoder inherits its target's spectrum. We propose LoCA (Location Contrastive Alignment), a two-stage pipeline that satisfies both. Stage 1 jointly trains Sentinel-1 and Sentinel-2 encoders with contrastive learning, producing a co-aligned, high-rank multimodal target. Stage 2 freezes the image encoders and trains a location and a tabular metadata encoder as projection adapters with multi-view contrastive learning. Against five state-of-the-art encoders on 17 tasks ranging from land cover and wildfire forecasting to geophysical, socio-economic, biodiversity and climate applications, LoCA achieves the most task wins, is the only model to outperform every other on a majority of tasks, and wins 68.8% of all per-task comparisons.
Figures & tables
Figure 1: LoCA pipeline. Stage 1 : jointly train Sentinel-1 SAR and Sentinel-2 optical encoders with DHEL to produce a co-aligned, high-rank multimodal target. Stage 2 : freeze the image encoders and fit a location and a metadata encoder as adapters against the four-view system. Inference: a coordinate query passes through the encoder alone to produce an embedding for downstream tasks.
Figure 2
Observation-based FMs
Location encoders
Task ↓ Model →
AlphaEarth
Copernicus FM
SatCLIP
GeoCLIP
GT-Loc
LoCA (Ours)
Metric
MLP
Linear
MLP
Linear
MLP
Linear
MLP
Linear
MLP
Linear
MLP
Linear
Biodiversity & Ecology
Countries
N/A
N/A
80.9
88.3
95.5
96.5
91.5
79.8
94.7
94.4
96.4
97.0
F1 w
Biome
N/A
N/A
84.0
82.4
89.8
86.6
86.9
75.6
88.1
81.9
91.0
87.0
F1 w
Ecoregions
N/A
N/A
59.9
75.1
79.7
84.4
60.6
42.6
69.6
75.9
80.2
85.0
F1 w
Table 2: Downstream evaluation. Mean over 10 seeds (3 for SatBird); metric ×100 , higher is better. Green : best across both probes; blue : best within probe (MLP, Linear) when not best overall. Environmental tasks predict relative ordering of climate variables. Bottom two rows: head-to-head comparisons ( Table 11 ). N/A: embeddings not retrievable ( Appendix E ).
Removed component
Biodiv. & Ecology
Land cover
Wildfire
Geophysical
Environmental
Stage 1
− 56.9
− 53.6
− 26.8
− 61.3
− 21.1
Metadata encoder
− 0.1
+ 0.5
+ 0.3
− 3.5
− 1.3
Table 3: Isolating the contributions of (i) the Stage 1 and (ii) the metadata encoder. We report the change in mean linear probe score per task family with respect to the full model. See also Table 13 .
Location
S1
S2
S1+S2
All
base
Stage 2
base
Stage 2
base
Stage 2
base
Stage 2
base
Stage 2
Mean F1
74.7
76.6
69.6
75.9
78.4
80.9
78.1
80.8
80.9
84.3
Table 4: Mean F1 on 10 single-label land-cover classes for 4 supervised encoders, before/after LoCA’s Stage 2 . “S1+S2”: trained on both modalities; “All”: concatenated embeddings.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Variable
Dimensions
Unit
Description
elevation
(lat, lon)
m
Height above reference ellipsoid
slope
(lat, lon)
rad
Magnitude of terrain gradient
aspect
(lat, lon)
deg
Direction of steepest ascent
curvature
(lat, lon)
–
Local topographic curvature
roads_dist_km
(lat, lon)
km
Distance to nearest OSM road
pop_density
(lat, lon, year)
persons/km 2
WorldPop annual estimates
Appendix
Table 5: Variables included in the global static-feature archive.
Hyperparameter
Stage 1
Stage 2
Modalities
S2, S1
static, location, S2, S1
Optimiser
SGD
SGD
Learning rate
0.2
0.0016
Momentum
0.9
0.9
Weight decay
1×10−6
5×10−6
Scheduler
warmup-cosine
warmup-cosine
Appendix
Table 6: Hyperparameters for the two-stage training pipeline. In both stages we use SGD with a linear warmup of 10 epochs followed by cosine annealing over the remaining epochs.
Task name
Source
Number of samples
Dataset split
Evaluation metric
Geographic coverage
Regression
Air temperature ( Hooker et al., 2018 )
SatCLIP
3076
60-20-20
R2
Global
Median income ( Jia and Benson, 2020 )
3219
50-20-30
R2
Cont. USA
Cali housing ( Pace and Barry, 1997 )
20640
50-20-30
R2
Calif., USA
Elevation ( Klemmer et al., 2025 )
100000
30-10-60
R2
Global
Population ( Klemmer et al., 2025 )
74512
30-10-60
R2
Global
Appendix
Table 7: Summary of benchmark evaluation tasks using the location encoder. For iNaturalist, 10% of the official training split is used for validation, and the official validation split is used as the test set.
Encoder
Embedding dimension
SatCLIP ( Klemmer et al., 2025 )
256
GeoCLIP ( Cepeda et al., 2023 )
512
GT-Loc ( Shatwell et al., 2025 )
512
AlphaEarth Foundations (AEF) ( Brown et al., 2025 )
64
Copernicus-FM ( Wang et al., 2025 )
768
LoCA (ours)
256
Appendix
Table 8: Dimensionality of the fixed representations extracted from each location encoder or geospatial foundation-model embedding product.
Hyperparameter
Value
Epochs
100
Batch size
1024
Optimizer
AdamW
Learning rate
10−3
Weight decay
10−5
MLP hidden dimension
128
Appendix
Table 9: Downstream training hyperparameters used for all linear and MLP downstream probes unless otherwise specified. The separate climate-ranking configuration is described in the text.
Task ↓ Model →
AlphaEarth
Copernicus FM
SatCLIP
Metric
MLP
Linear
MLP
Linear
MLP
Linear
Land cover
reBEN
42.8 ± 0.43
43.3 ± 0.71
57.4 ± 0.37
55.2 ± 0.28
57.3 ± 0.54
54.9 ± 0.59
F1 μ
Wildfire forecasting
Wildfire forecasting
80.5 ± 0.59
70.1 ± 0.43
74.9 ± 0.16
71.0 ± 0.27
79.1 ± 0.47
75.8 ± 0.31
AUPRC
Geophysical & Socio-economic
Appendix
Table 10: Extended results for Table 2 : mean ± std over 10 seeds (3 in SatBird), metric ×100 , higher is better. Both halves share one highlighting convention across all six models: best across both probes in green, best per probe in blue when not best overall, second per probe in bold . N/A indicates the embeddings could not be retrieved ( Appendix E ).
Row beats column →
LoCA
SatCLIP
AlphaEarth
GT-Loc
GeoCLIP
Copernicus FM
Majority wins
Win rate
LoCA (Ours)
–
12/17
5/9
13/17
12/17
11/17
5/5
68.8%
SatCLIP
5/17
–
5/9
11/17
10/17
12/17
4/5
55.8%
AlphaEarth
4/9
4/9
–
6/9
6/9
6/9
3/5
57.8%
GT-Loc
4/17
6/17
3/9
–
12/17
10/17
2/5
45.5%
GeoCLIP
5/17
7/17
3/9
4/17
–
9/17
1/5
36.4%
Copernicus FM
4/17
5/17
3/9
7/17
8/17
–
0/5
35.1%
Appendix
Table 11: Head-to-head comparison over the 17 tasks of Table 2 (best probe per model). Cell (i,j) : tasks on which the row model outperforms the column model, out of those both were evaluated on; green marks a majority. Ties at displayed precision count for neither model. LoCA is the only encoder that beats every other encoder on a majority of shared tasks.
Model
Mean F1
Arable Land
Mixed Forest
Coniferous Forest
Transitional Woodland
Broad-leaved Forest
Urban Fabric
Inland Waters
Complex Cultivation
Marine Waters
Pastures
Location
74.7
71.0
77.1
80.4
67.5
73.8
65.8
72.8
69.6
89.3
79.6
Location (ours)
76.6
71.4
79.8
83.8
68.5
77.6
63.0
73.0
70.7
97.8
80.6
S1
69.6
72.9
70.0
73.1
63.8
64.0
75.9
78.4
52.5
95.2
50.1
S1 (ours)
75.9
76.1
74.6
76.7
67.0
70.2
78.0
84.7
64.8
95.6
70.9
S2
78.4
78.5
78.4
84.5
68.9
76.7
73.8
82.5
67.9
96.1
76.6
S2 (ours)
80.9
81.7
81.7
86.9
71.4
79.1
75.1
83.5
71.4
96.7
81.2
Appendix
Table 12: Per-class F1 on BigEarthNet land-cover classification (10 classes), comparing pure location encoders (rows 1–2) against supervised ResNet-18 image encoders with and without the MV-DHEL multimodal training. “Mean F1” is the unweighted mean across the 10 class-wise F1 scores; remaining columns report class-wise F1. Best per column is highlighted in green and second best in bold .
Task ↓ Variant →
LoCA w/o Stage 1
LoCA w/o metadata
LoCA (full)
Metric
Biodiversity & Ecology
Countries
4.4
96.9
97.0
F1 w
Biome
9.7
85.8
87.0
F1 w
Ecoregions
1.7
85.4
85.0
F1 w
iNaturalist
56.6
71.6
71.4
Acc.
SatBird (USA summer)
44.5
61.2
61.1
top- k Acc.
Appendix
Table 13: Ablation isolating the contributions of (i) the joint Stage-1 contrastive pretraining and (ii) the metadata encoder, under the linear probe. Values are metric ×100 ; higher is better. Highlighting follows Table 2 . Without Stage 1 the embeddings collapse: downstream probes drop to (or below) the trivial predictor and environmental ranking is at chance ( 50 ). All variants are evaluated on all 17 tasks.
Asset
Use in this paper
License / terms
SatCLIP S2-100K dataset
Starting point for coordinate sampling and Sentinel-2 pairing.
Research purposes only; license field not specified in the dataset card.
SatCLIP code and pretrained models
Baseline location encoder; architecture reference.
MIT License.
Copernicus DEM GLO-30
Elevation, slope, aspect, and curvature metadata.
Copernicus DEM free license.
WorldPop population density
Population-density metadata.
Creative Commons Attribution 4.0 International (CC BY 4.0).
WorldPop distance-to-roads product
Road-distance metadata.
Creative Commons Attribution 4.0 International (CC BY 4.0).
OpenStreetMap
Underlying source for road-distance data.
Open Data Commons Open Database License 1.0 (ODbL 1.0).
Appendix
Table 14: Existing assets used in this work and their licenses or terms of use. “Use in this paper” describes how the asset is used; it does not imply that we redistribute the original asset.
The plethora of readily available geospatial data offers exciting opportunities to learn high quality representations of the planet, but the sheer size of the Earth Observations (EO), differing modalities, and different sensor types pose significant challenges in doing so. Location encoders have emerged as an efficient way of compressing EOs into location-specific embeddings. However, current state-of-the-art location encoders rely on computationally expensive CLIP-style frameworks that require large batch sizes in the 16K--32K range, suffer from false negative samples, and scale poorly with additional modalities. We introduce the Scalable Location Encoder via Distillation (SLED), a distillation-based location encoder that uses geospatial location as a binding modality to pretrain location encoders with any modality of geospatial data. The resulting location encoder framework is lightweight, modular, and can flexibly incorporate multiple modes, while eliminating the need for spatiotemporal coregistration of samples. SLED is performant with batch sizes as small as 128, enabling pretraining at a fraction of the runtime and compute costs of current state-of-the-art models. We demonstrate our approach by pretraining unimodal and multimodal SLED models on Sentinel-1, Sentinel-2, and Landsat imagery. We show that both unimodal and multimodal SLED models keep pace with or outperform existing approaches on a diverse set of 19 human-centric benchmark tasks and explore the benefits of using additional modes in pretraining.
Spatial prediction tasks are often limited by a lack of high-quality labelled ground-truth observations. To overcome this challenge, self-supervised pre-training is a possible solution, with contrastive learning dominant for location encoders. Those approaches usually align geographic coordinates with just one additional modality. We propose two multimodal contrastive learning architectures: Multimodal Embedding via Location Tying (MELT) and Sequential Alternating Location Training (SALT). These architectures expand this framework beyond two modalities by utilising unpaired geospatial data. Both methods are technically viable and match the performance of the strongest two-modality baseline (SATCLIP) across four downstream tasks. However, increasing the number of modalities does not consistently improve performance, suggesting that the chosen location encoder is the main limitation - the contrastive objective reaches its peak early, regardless of modality diversity or pre-training volume. MELT provides more stable training than SALT and presents a stronger foundation for future scaling.
Jonathan Hecht, Lukas Arzoumanidis, Ziyue Li +1
Computational Methods Lab, HafenCity University Hamburg, Germany. · Institute of Geodesy and Geoinformation, University of Bonn, Germany. · Dept. of Operations & Technology, Technical University of Munich; Heilbronn Data Science Center; Munich Data Science Institute.
The growing availability of co-located geospatial data spanning aerial imagery, street-level views, elevation models, text, and geographic coordinates offers a unique opportunity for multimodal representation learning. We introduce UNIGEOCLIP, a massively multimodal contrastive framework to jointly align five complementary geospatial modalities in a single unified embedding space. Unlike prior approaches that fuse modalities or rely on a central pivot representation, our method performs all-to-all contrastive alignment, enabling seamless comparison, retrieval, and reasoning across arbitrary combinations of modalities. We further propose a scaled latitude-longitude encoder that improves spatial representation by capturing multi-scale geographic structure. Extensive experiments across downstream geospatial tasks demonstrate that UNIGEOCLIP consistently outperforms single-modality contrastive models and coordinate-only baselines, highlighting the benefits of holistic multimodal geospatial alignment. A reference implementation is available at https://gastruc.github.io/unigeoclip.
Guillaume Astruc, Eduard Trulls, Jan Hosang +2
LASTIG, Univ Gustave Eiffel, IGN, ENSG, France · CNES, France · LIGM, CNRS, Univ Gustave Eiffel, ENPC, Institut Polytechnique de Paris, Marne-la-Vallée, France +1