Learning Administrative Boundary-Aware Location Encoders via Reverse Geocoding

Lakshay Sharma; Alex Marin

cs.LGOct 16, 2026

Learning Administrative Boundary-Aware Location Encoders via Reverse Geocoding

Authors: Lakshay Sharma, Alex Marin

Organizations: Instacart New York, NY, USA · Thomson Reuters Labs Seattle, WA, USA

Abstract

Location encoders learn compact representations of latitude--longitude coordinates for downstream geospatial prediction tasks. Existing approaches are pretrained using continuous or naturally occurring supervision, such as imagery, climate variables, or species distributions, and therefore do not explicitly capture administrative boundaries or place names, despite their strong influence on many real-world geospatial phenomena. We introduce, to our knowledge, the first method for learning administrative boundary-aware location encoders, using reverse-geocoded text as a self-supervised training signal. Our approach can be applied to existing location encoder architectures and consistently improves their downstream performance. We further introduce RevGeoRQ, a custom architecture tailored to reverse-geocoding supervision, which yields additional gains and exceeds contemporary location encoders on several downstream tasks. Finally, the use of human-interpretable reverse-geocoded strings and residual-quantized semantic IDs provides a more interpretable representation of geographic location than existing approaches. Code and model weights to be released soon.

Figures & tables

Figure 1: Left: Pretraining architecture. Right: Learned $q^{(0)}$ (level-0) codes from RQVAE Semantic IDs on S2-100K points

	Regression					Classification
Model	air_temp	elevation	population	cali_housing	avg. [2pt]R 2	country	eco regions	biome	avg. [2pt]acc.
Baseline
Qwen0.6B	$.820\pm.005$	$.727\pm.006$	$.692\pm.013$	$.675\pm.003$	$.729$	$.926\pm.003$	$.489\pm.007$	$.705\pm.009$	$.707$
SatCLIP
Native	$.890\pm.008$	$.902\pm.006$	$.812\pm.003$	$.718\pm.001$	$.831$	$.907\pm.009$	$.689\pm.010$	$.841\pm.005$	$.812$
TextRQ	$\underline{.906\pm.009}$	$\underline{.903\pm.005}$	$\underline{.825\pm.010}$	$.687\pm.002$	$.830$	$\underline{.946\pm.002}$	$\underline{.719\pm.013}$	$\underline{.849\pm.012}$	$\underline{.838}$

Table 1: Results across downstream datasets. Bold marks the best mean per dataset; underline indicates a TextRQ or Combined variant that outperforms its corresponding Native model. Regression tasks report

R^{2}

; classification tasks report accuracy.

Model	top-1	top-3	top-5	mean
Baseline
Vision-only (Inception-v3)	$0.6617$	$0.8331$	$0.8803$	$0.7917$
SatCLIP
Native (L=40)	$0.7537\pm 0.0009$	$0.8945\pm 0.0003$	$0.9280\pm 0.0003$	$0.8587$
TextRQ (L=10)	$\underline{\mathbf{0.7730\pm 0.0008}}$	$\underline{\mathbf{0.9019\pm 0.0006}}$	$\underline{\mathbf{0.9306\pm 0.0004}}$	$\underline{\mathbf{0.8685}}$
Combined	$\underline{0.7677\pm 0.0007}$	$\underline{0.8964\pm 0.0006}$	$0.9256\pm 0.0003$	$\underline{0.8633}$

Table 2: iNaturalist 2018 geo-prior evaluation. Values report top-

k

validation accuracy as mean

\pm

std over three seeds; the mean column gives the per-row average over top-1, top-3, and top-5. Bold marks the best mean within each model family; underline indicates improvement over the corresponding Native baseline.

	Regression					Classification
Model	air_temp	elevation	population	cali_housing	avg. [2pt]R 2	country	eco regions	biome	avg. [2pt]acc.
Qwen-0.6B	$.820\pm.005$	$.727\pm.006$	$.692\pm.013$	$.675\pm.003$	$.729$	$.926\pm.003$	$.489\pm.007$	$.705\pm.009$	$.707$
RQVAE-only (0.6B)	$.913\pm.005$	$.564\pm.165$	$.683\pm.049$	$.508\pm.021$	$.667$	$.907\pm.010$	$.560\pm.022$	$.771\pm.005$	$.746$
SatCLIP-TextRQ (L=40, 0.6B)	$-.028\pm.019$	$.649\pm.019$	$.458\pm.038$	$\mathbf{.729\pm.013}$	$.452$	$.382\pm.012$	$.266\pm.003$	$.415\pm.008$	$.354$
SatCLIP-TextRQ (L=10, 0.6B)	$.906\pm.009$	$.903\pm.005$	$.825\pm.010$	$.687\pm.002$	$.830$	$.946\pm.002$	$.719\pm.013$	$.849\pm.012$	$.838$
SatCLIP-TextRQ (L=10, 4B)	$.922\pm.005$	$.890\pm.011$	$.831\pm.001$	$.682\pm.019$	$.831$	$.949\pm.002$	$.725\pm.005$	$.853\pm.009$	$.842$

Table 3: Model architecture and parameter ablations. Regression reports

R^{2}

, classification reports accuracy, and avg. denotes the row mean within each group. Bold marks the best value per column.

Figure 2: Left: the reverse-geocoded continent share within each $q^{(0)}$ code. Center: the eight largest $q^{(1)}$ codes within $q^{(0)}=4$ . Right: attribute accuracy vs. cumulative prefix depth.

Figure 3: Cosine similarity of each land embedding to a query point ( $\star$ ) — a Himalayan query ( top ) and an Atacama query ( bottom ) — for our TextRQ encoder (left) vs. native SatCLIP L=10 and L=40 (center/right).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

	SatCLIP	GeoCLIP	Climplicit	SINR
Backbone variant	SatCLIP ViT16-L10	GeoCLIP	Climplicit	$\mathcal{L}_{\text{AN-full}}$ SINR (Coords., 1000/class)
Training data	S2-100K	MP-16 Flickr	CHELSA land grid	iNaturalist
(# points)	(91,858)	( $\sim$ 4.72M)	( $\sim$ 224M)	( $\sim$ 35.5M)
Budget	max_epochs =100	max_steps =50k	max_steps =50k	max_epochs =50
Optimizer	AdamW (betas $0.9,0.999$ ; eps $10^{-8}$ )
LR	peak $3\times 10^{-4}$ , weight decay $10^{-4}$

Appendix

Table 4: Per-encoder pretraining setup

	Regression					Classification
Model [2pt] (all TextRQ)	air_temp	elevation	population	cali [2pt]housing	avg. [2pt]R 2	country	eco [2pt]regions	biome	avg. [2pt]acc.
SatCLIP (L=10)
CLIP	$\mathbf{.930\pm.009}$	$.888\pm.007$	$\mathbf{.831\pm.002}$	$.671\pm.006$	$\mathbf{.830}$	$\mathbf{.948\pm.004}$	$\mathbf{.723\pm.009}$	$\mathbf{.859\pm.007}$	$\mathbf{.843}$
RQVAE	$.906\pm.009$	$\mathbf{.903\pm.005}$	$.825\pm.010$	$\mathbf{.687\pm.002}$	$\mathbf{.830}$	$.946\pm.002$	$.719\pm.013$	$.849\pm.012$	$.838$
GeoCLIP
CLIP	$.754\pm.010$	$.436\pm.044$	$.253\pm.027$	$.694\pm.007$	$.534$	$.767\pm.003$	$.414\pm.001$	$.651\pm.004$	$.611$

Appendix

Table 5: Downstream performance after ablating the pretraining objective on each encoder’s own data. Bold shows best per backbone-task pair between CLIP and RQVAE objectives

Figure 4: Cell spread per level (cumulative vs. marginal).

Model (dataset)	Total points	Distinct points	Distinct reverse [2pt]geocoded strings	Reduction rate [2pt](distinct points per string)
SatCLIP (S2-100K)	91,858	91,858	21,188	4.3 $\times$
GeoCLIP (MP-16)	4,723,695	2,875,487	81,997	35.1 $\times$
Climplicit (CHELSA)	224,368,589	224,368,589	94,473	2,375.0 $\times$
SINR (iNat)	35,500,262	27,862,308	101,411	274.7 $\times$

Appendix

Table 6: Input-space reduction from reverse geocoding. The reduction rate is the number of distinct points divided by the number of distinct reverse geocoded strings, i.e. how many distinct input locations collapse onto each text target.

Geospatial Representation Learning

Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.

See all 128 papers

Do Location Encoders Capture Spatial Effects? A GeoShapley Benchmark Across Scales
Jun 22, 2026Daniel Kiv, Shaowen WangGeospatial Representation Learning Shapley Value Attribution
Siebel School of Computing & Data Science, University of Illinois Urbana-Champaign Urbana, IL, USA · Department of Geography & GIS and Siebel School of Computing & Data Science, University of Illinois Urbana-Champaign Urbana, IL, USA
ReaGeo: Reasoning-Enhanced End-to-End Geocoding with LLMs
Apr 23, 2026Jian Cui, Zhiyuan Ren, Desheng Weng +4Geospatial Reasoning
Amap, Alibaba Group · Tsinghua University
SLED: Scalable Location Encoding via Distillation
Aug 6, 2026Kevin Lane, Zhongying Wang, Esther Rolf +1Geospatial Representation Learning Multimodal Pretraining
University of Colorado Boulder

Jun 22, 2026J/K move · Enter open · S save

Do Location Encoders Capture Spatial Effects? A GeoShapley Benchmark Across Scales

Daniel Kiv, Shaowen Wang

Siebel School of Computing & Data Science, University of Illinois Urbana-Champaign Urbana, IL, USA · Department of Geography & GIS and Siebel School of Computing & Data Science, University of Illinois Urbana-Champaign Urbana, IL, USA

Geospatial Representation Learning Shapley Value Attribution

Open paper AskPDF arXiv

Location encoders transform geographic coordinates into high dimensional embeddings for downstream machine learning, but it is unclear how well these representations capture interpretable spatial effects. We benchmark whether GeoShapley, a game-theoretic explainer that treats all location features as a single joint player, can recover spatially varying coefficients from models built on location-encoder embeddings. Eleven encoders from the TorchSpatial framework are evaluated against a synthetic process with known coefficients, across three scales (grid, county, global), with and without raw coordinates alongside the embedding, and under untrained and contrastively trained conditions. Measuring recovery as the correlation between estimated and true coefficients, we report how it varies with scale and encoder architecture and compare the embeddings against a raw-coordinate baseline. Recovery of the primary coefficient is consistently high across encoders, whereas recovery of a secondary coefficient is more scale-dependent, differing most at the global scale; the raw-coordinate baseline remains competitive throughout.

Abstract

Geospatial Representation Learning

Explore similar work

Do Location Encoders Capture Spatial Effects? A GeoShapley Benchmark Across Scales

ReaGeo: Reasoning-Enhanced End-to-End Geocoding with LLMs

SLED: Scalable Location Encoding via Distillation