acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Urban Representation Pretraining from Large-Scale Geospatial Data

31 views

Abstract

This paper studies how to encode cities with a Transformer for urban representation pretraining on large-scale geospatial data. Geospatial entities form the urban fabric of blocks, which in turn compose larger regions that collectively constitute cities. This entity–block–region–city hierarchy provides a natural basis for urban representation learning. Scaling it across thousands of cities requires modeling local functional interactions within blocks and spatial organization within and across regions. To this end, we propose HESTIA, a hierarchical framework for urban representation pretraining. Specifically, an Urban Fabric Tokenizer converts heterogeneous geospatial data into entity-category tokens arranged in a unified tensor reflecting the native city hierarchy. The Urban Interaction Transformer (UiT) applies attention in three nested stages, allowing urban semantics to emerge progressively through the hierarchy: first from functional interactions among entity categories within blocks, then from geometry-aware interactions among blocks within regions, and finally from city-scale contextualization across regions. Pretrained on a large-scale U.S. geospatial dataset of 8.18 million blocks and over 263 million entities, HESTIA achieves state-of-the-art performance on the multilevel benchmark, outperforming strong baselines including DeepMind's AlphaEarth, validating its hierarchical pretraining across spatial scales. Code is available at https://anonymous.4open.science/r/HESTIA-0D1C/.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Urban computing, urban region representation learning, urban representation learning, geospatial representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QJtqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HESTIA, a hierarchical urban representation pretraining framework that organizes large-scale geospatial data into an entity–block–region–city hierarchy. It introduces an Urban Fabric Tokenizer and an Urban Interaction Transformer (UiT) with three-stage attention, trained through local function and regional fabric reconstruction. Experiments on nationwide U.S. geospatial data demonstrate strong performance across multiple spatial scales.

Critical strengths

S1: The framework is pretrained on over 263 million entities and 8.18 million blocks, demonstrating scalability.

S2: Extensive comparisons with 21 baselines, including AlphaEarth and S2Vec, show strong performance across entity-, block-, and region-level tasks.

S3: The hierarchical tokenization and attention design achieves competitive performance with only 12.9M parameters.

Critical weaknesses

Limited methodological novelty: The hierarchical attention architecture largely combines existing Transformer mechanisms, category aggregation, positional encoding, and masked reconstruction objectives. The technical contributions are primarily architectural integration rather than fundamentally new learning mechanisms.

Insufficient isolation of architectural contributions: Although component ablations are provided, the paper lacks a direct comparison with simpler hierarchical pooling or hierarchical Transformer baselines under matched inputs and training objectives, making it difficult to attribute the improvements specifically to UiT.

Limited geographic generalization: Pretraining and evaluation focus exclusively on U.S. geospatial data. Generalization across countries with different urban structures and data distributions remains unverified.

Information loss during tokenization: Aggregating entities of the same category into a single token discards fine-grained spatial arrangements and individual entity relationships, potentially limiting representation expressiveness.

Questions for authors

See the above

8 comments
AuthorsOpenReview ↗

Response Overview

We thank the reviewer for recognizing HESTIA's data scale, geographic coverage, and overall performance. To be honest, we initially felt sad and discouraged when we first read the review. As we worked through each comment, however, we gradually realized that these questions were not simply a dismissal of our work. We must say that you gave the paper a renewed sense of purpose.

Over the past ten days, we invested substantial time in additional experiments and careful result verification. The most important and time-consuming addition was to use unaggregated individual entities as the inputs to entity attention, replacing the exclusive use of the Entity-category tokens in the original paper. We also completed experiments using H3 hexagons as alternative spatial units and cross-country generalization evaluations in Canada and Mexico with the frozen U.S.-trained checkpoint. This work required downloading large-scale map data, map projection, spatial indexing, label alignment, and careful alignment with AlphaEarth embeddings. Completing the full evaluation therefore required substantial time for implementation, computation, and verification. We sincerely apologize for keeping you waiting and greatly appreciate your patience while we completed this work. All of these experiments have now been completed and incorporated into a unified evaluation, and we will release the corresponding benchmark to support future research in the urban computing community. On this basis, we compared against S2Vec and AlphaEarth using the same spatial folds and probe protocol.

Below, we respond to the weaknesses in reverse order. First, for Weakness 4, we discuss the loss of fine-grained spatial information caused by tokenization and compare Entity-category tokens with individual-entity tokens at the region, block, and entity levels. Second, for Weakness 3, we evaluate cross-country and cross-unit generalization by applying the frozen U.S.-trained checkpoint in Canada and Mexico and by replacing Census units with H3 hexagons. Third, for Weakness 2, we add Hierarchical pooling and Hierarchical Transformer controls using the same reconstruction objectives and downstream protocol. Finally, for Weakness 1, we clarify that HESTIA contributes a unified and scalable framework for hierarchical urban pretraining.

AuthorsOpenReview ↗

Rebuttal to Weaknesses 4, Part 1

[Weakness 4, Part 1: Individual-Entity Tokenization and Region-Level Evaluation]

This is one of the issues we reflected on most deeply during this review and one we especially wanted to address carefully. The original design aggregates entities of the same category within a block into one token, so it no longer retains the independent coordinates of each entity or the relationships among individual entities. This choice does not reflect a belief that fine-grained spatial relationships are unimportant. It is a trade-off between computational scalability at the national scale and spatial detail. After aggregation, the model still retains the entity category, count or coverage area, block area, and the relative positions of blocks and regions. It can therefore still model local functional composition and spatial structure across blocks. More importantly, computation grows primarily with the number of categories actually present in each block, not with the raw entities, which reduces the computational overhead of nationwide pretraining.

The reviewer's comment also prompted us to ask a further question: what happens if entities of the same category are no longer aggregated and the model instead performs attention directly among individual entities? Can the model retain the spatial relationships removed by aggregation? Frankly, this was also the direction we were most eager to explore after reading the review. Over the past ten days, we trained and carefully checked multiple complete pretraining variants. Once the tokens changed from aggregated categories to individual entities, the sequences became substantially longer, increasing both memory consumption and training time. Each complete variant required approximately 8--10 hours. This was the most resource-intensive addition in this review cycle and the main reason our response took longer than expected. Fortunately, the effort was worthwhile. It clarified both the limits of category aggregation and the settings in which individual-entity modeling is most valuable.

In the new design, each entity is represented by an individual entity token encoding its category, relative (x,y)(x,y) coordinates and coverage area. All entity tokens use two-dimensional positional encoding at the entity level. The central change from the original design is that attention now operates on individual entity tokens instead of aggregated entity-category tokens. The model can therefore learn relationships directly among entities, not only interactions among categories. For clarity, we refer to the new design as individual-entity tokens and the original design in the paper as entity-category tokens.

We first examined whether the finer-grained individual-entity input also improves representations at aggregated spatial levels. For a fair comparison, the entity-category-token and individual-entity-token variants use the same downstream tasks, seven spatial folds, and MLP probe. The results for the original method, AlphaEarth, and S2Vec are taken directly from the formal evaluation in the paper. Land use uses Macro-AUPRC, Age distribution uses KL divergence, and all other tasks use R2R^2. Lower is better for Age distribution, while higher is better for all other tasks. The region-level results are as follows:

Region-level taskEntity-category tokens (original method)Individual-entity tokensAlphaEarthS2Vec
Land use ↑0.5545 ±\pm 0.01990.5590 ±\pm 0.01730.4650 ±\pm 0.02340.4872 ±\pm 0.0242
Road density ↑0.8956 ±\pm 0.02530.8921 ±\pm 0.01970.8556 ±\pm 0.02520.7726 ±\pm 0.0296
Population density ↑0.7431 ±\pm 0.07420.7736 ±\pm 0.05490.5765 ±\pm 0.11210.5340 ±\pm 0.0800
Age distribution ↓0.0662 ±\pm 0.00730.0671 ±\pm 0.00730.0958 ±\pm 0.01520.0742 ±\pm 0.0069
Nightlight ↑0.8356 ±\pm 0.02350.8071 ±\pm 0.03570.7469 ±\pm 0.08590.7401 ±\pm 0.0677
Household income ↑0.6023 ±\pm 0.05680.5951 ±\pm 0.05450.1202 ±\pm 0.26090.5286 ±\pm 0.0475
CO$_2$ ↑0.2988 ±\pm 0.08730.3093 ±\pm 0.12640.2117 ±\pm 0.18030.2613 ±\pm 0.1228
Urban heat ↑0.5635 ±\pm 0.05830.5471 ±\pm 0.05320.5612 ±\pm 0.09310.3778 ±\pm 0.0699

Part 2 continues this comparison at the block and entity levels.

AuthorsOpenReview ↗

Rebuttal to Weaknesses 4, Part 2

[Weakness 4, Part 2: Block- and Entity-Level Evaluation]

Continuing the same comparison and evaluation protocol described in Part 1, we evaluated the block-level representations:

Block-level taskEntity-category tokens (original method)Individual-entity tokensAlphaEarthS2Vec
Land use ↑0.4135 ±\pm 0.01740.4161 ±\pm 0.01670.2622 ±\pm 0.01430.1939 ±\pm 0.0108
Population log1p ↑0.6205 ±\pm 0.02060.6093 ±\pm 0.02110.4059 ±\pm 0.03970.2399 ±\pm 0.0543

These results do not support the simple conclusion that finer tokens always perform better. At the region and block levels, individual-entity tokens perform better on three of the eight region-level tasks and slightly better on one of the two block-level tasks. Entity-category tokens perform better on the remaining tasks. However, the better result of the two HESTIA tokenizations exceeds both AlphaEarth and S2Vec on all ten region-level and block-level tasks. Although the outcome is not uniform, it clarifies the respective roles of the two designs. Adding object-level detail does not automatically improve tasks that depend mainly on the aggregate composition of a region.

This is not, however, the most important part of the experiment. The information loss raised by the reviewer occurs directly at the entity level. We therefore added a unified entity-category completion experiment that directly tests whether the model can use fine-grained entity context to recover a masked category. The two models use the same samples, seven spatial folds, and MLP probe. The category information of the target entity is removed before encoding. The results are as follows:

Entity-level representationTop-1 (%) ↑Top-5 (%) ↑
Entity-category tokens (paper checkpoint, reevaluated under the same protocol)32.7 ±\pm 0.863.5 ±\pm 0.8
Individual-entity tokens39.1 ±\pm 0.968.6 ±\pm 1.0

Under this unified protocol, individual-entity tokens improve Top-1 and Top-5 by 6.4 and 5.1 percentage points, respectively. This shows that the most direct value of individual-entity input appears precisely in tasks that genuinely depend on object-level information. It does not cause every aggregated-scale task to improve simultaneously. In other words, entity-category tokens still provide an efficient representation suitable for nationwide pretraining, while individual-entity tokens further enable the model to express entity-level spatial relationships.

Frankly, after completing these experiments, we were convinced by the individual-entity-token direction suggested by the reviewer. It not only produces clearer gains at the entity level but also reduces predefined category aggregation, giving the model an opportunity to learn structure directly from raw entities and their spatial relationships. This observation also echoes The Bitter Lesson: scalable learning methods should be allowed to discover structure directly from data and computation without excessive hand-designed abstractions. In the revised manuscript, we will therefore use individual entities as the objects of entity-level attention and fully report the experiments above, their computational cost, and the conditions under which each design is most appropriate. The reviewer's comment did more than prompt an additional experiment. It changed how we understand the boundary of the method and its next direction, making the paper more complete.

We once again thank you for your constructive feedback.

Reviewer_HdJFOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HESTIA for ubran representation learning from large-scale geospatial data. It represents a city through a hierarchy of entities, blocks, and regions. A Transformer then models interactions at each level in turn. The model is pretrained with two tasks, i.e., predicting masked categories in a block and reconstructing the category values of a region. HESTIA is pretrained on a large scale of geospatial and census data from US and evaluated on a variety of tasks at different spatial granularities. Overall, it shows strong empirical results and outperforms baseline methods.

Critical strengths

S1. This is a strong engineering paper, and the experiment is comprehensive. Models are pretrained on a sufficiently large dataset. S2. The method is well designed. The proposed three attention stages follow the natural structure of urban areas. S3. The ablation results can confirm both pretraining objectives help the region tasks and the authors also report training cost, which is helpful.

Critical weaknesses

W1. The current comparisons do not fully separate the effect of the model from the effect of its inputs. The baselines use different types of data, and some use released weights while others are pretrained on the national corpus. Table 1 shows large improvements on several tasks, but I am not sure whether some of these gains come from features that are strongly related to the downstream labels. A simple model using the same input data, block geometry, and surrounding blocks would make the comparison stronger.

W2. The two pretraining objectives have clear support from the ablations. However, the benefit of some architecture choices is less clear. Per Tables 18 and 19, could you explain which components are important for the whole method and which are useful only for some tasks and why?

W3. As I noted in the strengths, I think this is a strong engineering paper, and the experiments are clearly comprehensive. However, I suppose the methodological novelty somewhat limited. It uses a similar hierarchical approach as related literatures, often with attention and interaction mechanisms.

Questions for authors

W1 and W2

5 comments
AuthorsOpenReview ↗

Response Overview

We thank the reviewer for the positive assessment of our work on HESTIA and for clearly identifying the need to further distinguish input factors, architectural components, and methodological contributions. To address these concerns carefully, we conducted several targeted analyses: we added a raw-input descriptor control using the same inputs as HESTIA, reorganized the roles of individual components by downstream task and hierarchy level, and revisited the methodological positioning of UiT relative to existing hierarchical Transformers. These additions now allow us to explain more clearly which gains arise from the inputs, which come from the model, and what this work ultimately aims to contribute.

We respond to the three weaknesses in turn. In Weakness 1, we distinguish the contribution of the input information from that of the model itself. In Weakness 2, we explain which tasks and hierarchy levels benefit from each architectural component. In Weakness 3, we discuss the methodological contribution of UiT as a unified framework for urban pretraining.

AuthorsOpenReview ↗

Rebuttal to Weaknesses 1

[Weakness 1: Effects of the Model and the Inputs]

We agree that the comparison should separate category composition, block geometry, and surrounding context. We therefore decompose the same map inputs into four fixed descriptors, all built for the same target regions and spatial windows. Category composition contains log-transformed POI category counts and AOI shares. Block geometry contains tract/block geometry and relative coordinates. Composition + block geometry concatenates these two views. Composition + block geometry + neighbor context additionally appends mean composition and geometry features from boundary-adjacent tracts in the observed adjacency graph, providing a controlled summary of surrounding blocks. The table reports eight representative tasks with the same seven spatial folds, common samples, hidden-1024 MLP probe, and evaluation metrics as the paper. Age distribution is lower-is-better. All other columns are higher-is-better.

MethodLand use ↑Road density ↑PM$_{2.5}$ ↑Age distribution ↓Nightlight ↑Household income ↑CO$_2$ ↑Urban heat ↑
Category composition only0.3760.5180.2920.0700.6120.4590.1740.430
Block geometry only0.2180.6830.2120.0940.4470.0470.1280.316
Composition + block geometry0.3860.6900.3060.0690.6580.4790.1850.437
Composition + block geometry + neighbor context0.3850.7910.3060.0670.6740.6280.2240.502
HESTIA0.5560.8950.4590.0660.8260.6130.2930.564

The table reveals a clear progression. Category composition already carries substantial signal, while geometry provides complementary information. Combining them improves all eight displayed tasks over either single-input descriptor. Adding neighbor context further improves six tasks and leaves PM$_{2.5}$ effectively tied at the shown precision. Land use is also essentially unchanged. HESTIA exceeds the full raw descriptor on seven of the eight tasks and remains close on household income, where the context descriptor is strongest. These controls therefore separate the signal available from each input family from the additional benefit of UiT's learned cross-level interactions and pretraining.

AuthorsOpenReview ↗

Rebuttal to Weaknesses 2

[Weakness 2: Scope of the Architectural Components]

To highlight the correspondence between components and levels, the table below lists tasks at the level on which each component acts most directly. Each value is the paired difference between the full model and the model with that component removed. A positive value indicates that the full model performs better.

ComponentPrimary levelCorresponding downstream taskFull-model referencePaired difference after removing the component
Block spatial attentionblockBlock population R2R^2 ↑0.618+0.193
Block spatial attentionregionHousehold income R2R^2 ↑0.603+0.102
Category objectiveblockBlock population R2R^2 ↑0.618+0.062
Category objectiveregionPM$_{2.5}$ R2R^2 ↑0.391+0.034
Regional objectiveregionHousehold income R2R^2 ↑0.603+0.066
Regional objectiveregionPopulation density R2R^2 ↑0.792+0.040

Block spatial attention yields the largest gains on block population and region-level household income, showing that it captures both spatial organization within blocks and cross-block context. The category objective supplements local category-composition information and contributes positively to tasks such as block population and PM$_{2.5}$. The regional objective mainly affects the overall structure at the region level, especially household income and population density.

Entity--category attention uses a different entity-level category-completion protocol, so it is not included in the region/block paired-difference table above. In the entity-level evaluation reported in the paper, HESTIA achieves Top-1 and Top-5 scores of 32.7% and 63.5%, respectively, compared with 25.5% and 56.1% for the raw block-composition control. These results are consistent with the role of this component in entity-category context modeling. Overall, the paired results reveal a division of roles aligned with the hierarchy and support the use of different UiT modules for context interaction, local category composition, region-level fabric, and entity-category completion.

Reviewer_WKK1OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduced HESTIA which is a hierarchical urban representation pretraining framework that aggregates geospatial entities into category tokens and applies attention across entity categories, blocks, and regions. It learns through local-function and regional-fabric reconstruction on a nationwide U.S. corpus, with evaluations at the entity, block, and region levels.

Critical strengths
  1. The category-based tokenization and hierarchical attention provide a coherent design for processing irregular geospatial data at national scale.

  2. The method achieves improvements on several downstream tasks.

  3. Geographically disjoint splits, unseen-county evaluation, and extensive structural ablations strengthen the experimental evidence.

Critical weaknesses
  1. Insufficient justification for the computational investment. The practical motivation for the Transformer architecture remains insufficiently convincing relative to its training cost. Its reported setup uses seven A100 GPUs, substantially more than earlier city-level methods, which raises accessibility concerns for public-sector users with limited computing resources. The paper should more clearly articulate what practical benefits justify this investment beyond improved downstream scores.

  2. Limited comparability with established urban-region benchmarks. Prior urban-region studies commonly evaluate regions within individual cities, whereas HESTIA adopts a nationwide tract–county hierarchy. Although this construction is motivated, the paper neither evaluates on established city-level benchmarks nor tests alternative spatial partitions, leaving unclear how robust its advantages are to the choice of geographic units and context.

  3. How are the padded dimensions (T), (B), and (R) determined for each batch, and are any hard caps imposed? Please report their observed ranges and maxima, specify how neighboring blocks are selected, and explain how oversized regions or counties are handled. If truncation, sampling, or chunking is used, how does it affect downstream performance and computational cost?

Questions for authors

See weakness.

7 comments
AuthorsOpenReview ↗

Response Overview

We sincerely thank the reviewer for the positive assessment of our paper and for emphasizing that computational accessibility and city-level comparability are just as important as downstream performance for practical deployment. These comments prompted us to think more carefully about how HESTIA can be evaluated and used beyond our original nationwide setting. We are sorry to have kept the reviewer waiting, and we deeply appreciate the reviewer's patience while we worked through the requested experiments as carefully as possible.

Over the past ten days, we devoted substantial effort to completing aggregate evaluations across ten U.S. cities, experiments using H3 hexagons as alternative spatial units, and frozen-transfer evaluations in Canada and Mexico using the U.S.-trained checkpoint. This involved far more than rerunning an existing evaluation. We downloaded and processed large-scale map data and the corresponding AlphaEarth embeddings, collected and cleaned socioeconomic and environmental indicators from different national sources, projected and aligned map entities and labels to official geographic units in Canada and Mexico and to U.S. H3 hexagons, constructed spatial indices, and repeatedly verified identifiers, labels, and evaluation folds. The large data volume and the differences among national data definitions made this a particularly labor-intensive and time-consuming process. We sincerely apologize again for the long delay. All of these experiments have now been completed and incorporated into a unified evaluation. We will also release the resulting benchmark to support future research in the urban computing community.

Our response is organized into three parts. In Weakness 1, we report the computational cost under single- and multi-GPU settings and explain the practical benefits that motivate this investment. In Weakness 2, we present city-level comparisons, evaluate an alternative H3 spatial partition, and test geographic transfer to Canada and Mexico. In Weakness 3, we clarify how batch dimensions and padding are determined, how neighboring regions are selected, and how large spatial units are divided and processed.

AuthorsOpenReview ↗

Rebuttal to Weaknesses 1

[Weakness 1-1: Computational Cost]

Regarding hardware requirements, we added a direct comparison between single-GPU and seven-GPU training. The seven GPUs are used primarily to reduce end-to-end waiting time. Seven GPUs are not required for training. Urban representation pretraining must process large-scale geospatial data, so approximately 30 GB of memory per GPU is an expected configuration. The two settings use the same training data, number of epochs, and model configuration. The seven-GPU setting processes seven mini-batches in parallel, with 16 regions per GPU, whereas the single-GPU setting processes one of these mini-batches at a time and accumulates gradients over multiple mini-batches before updating the parameters. The two settings are therefore directly comparable.

SettingNumber of GPUsPer-GPU batch sizeActual runtime (start to finish)Peak memory per GPU
Seven-GPU distributed training7167.54 hours26.94 GiB
Single-GPU control11622.91 hours28.73 GiB

The batch size in the table denotes the number of regions sent to the model on each GPU at one time. “Actual runtime” is the elapsed time from the start to the completion of training. Seven GPUs reduce the waiting time from approximately 22.9 hours to approximately 7.5 hours. The single-GPU peak memory is approximately 28.7 GiB, showing that training can also be completed on one GPU. For a GPU with less memory, the batch size can be reduced, i.e., fewer samples can be sent to the model at a time. This reduces processing speed but still allows training to be completed. Resource-constrained users can also directly use the released checkpoint or embeddings without repeating pretraining.

[Weakness 1-2: Practical Benefits]

Unified architecture. Transformers have become a general-purpose architecture for large-scale pretraining in modalities such as images, text, and video, and urban data can likewise be organized as hierarchical inputs that support batched computation. We organize the entity--block--region--city hierarchy within a single Urban Interaction Transformer (UiT), which learns cross-level relationships through unified hierarchical attention and masked reconstruction objectives. Existing urban studies also employ Transformers or interaction modules for particular functions, usually in combination with other modules. In contrast, UiT places hierarchical inputs, cross-level interactions, and reconstruction objectives within a single data flow, providing a more unified, simpler, and more extensible backbone for urban representation pretraining.

Practical application value. Open maps such as OpenStreetMap and Overture Maps often contain missing entity categories or incomplete annotations, and completing them one by one through manual annotation is costly. UiT's entity-category reconstruction head can use the entity context within the same block to provide category candidates for unlabeled entities. In addition to supporting downstream tasks in morphology, demographics, economy, and environment, the block and region representations can be used for urban-unit similarity retrieval, map-quality screening, and prioritization for manual review. The model outputs can therefore assist manual annotation and map quality control without replacing expert verification.

From single-city models to a reusable backbone. Many existing methods are trained primarily within a single city and require a separate set of parameters to be learned for each city. Serving 1,000 cities would therefore require repeated training or the maintenance of many city-specific models. HESTIA aims to pretrain a reusable nationwide backbone that uses the same set of parameters to generate representations for multiple cities and new spatial units. Downstream tasks need only train a lightweight probe and do not need to relearn a complete spatial encoder for every city. Please refer to our response to [Weakness 2-3: Generalization across Geographic Scopes] below for further details.

AuthorsOpenReview ↗

Rebuttal to Weaknesses 2

[Weakness 2-1: City-Level Comparability Evaluation]

Nationwide benchmark. Existing methods typically evaluate at most four cities without unified nationwide inputs or splits. Our U.S. benchmark covers 84,414 tracts, 8,180,866 blocks, and 12 tasks, with additional evaluation in Canada and Mexico. We will release all resources. Counties define national splits, and a fixed county--city map supports city-level reporting.

City-level results. We compare HESTIA with S2Vec and AlphaEarth on four tasks in ten U.S. cities. Cells report HESTIA/AlphaEarth/S2Vec. Lower is better only for Age KL.

CityLand use ↑Road density ↑Population density ↑Age distribution ↓
Los Angeles0.484/0.399/0.3680.633/0.264/0.2220.666/0.363/0.2550.058/0.081/0.066
New York0.445/0.291/0.3170.204/0.121/-0.3010.107/0.116/-0.0610.087/0.165/0.116
Chicago0.531/0.442/0.474-0.116/0.527/0.0680.720/0.398/0.3800.067/0.084/0.072
Houston0.482/0.449/0.4330.786/0.624/0.5390.599/-0.338/0.2750.070/0.118/0.079
Philadelphia0.565/0.474/0.4520.762/0.303/0.0760.672/0.300/0.1080.093/0.126/0.113
Dallas0.491/0.423/0.4290.756/0.622/0.4920.538/0.008/0.1490.073/0.125/0.082
Seattle0.616/0.524/0.5180.841/0.814/0.6190.871/0.598/0.5760.060/0.089/0.068
Boston0.435/0.390/0.3660.538/0.262/0.1990.563/0.451/0.1580.117/0.180/0.149
San Francisco0.510/0.406/0.3810.330/0.172/-0.2150.592/-0.392/0.0740.106/0.121/0.109
Washington DC0.451/0.435/0.3670.649/0.509/0.0650.687/0.337/-0.4250.133/0.145/0.146

[Weakness 2-2: Alternative Geospatial Partition]

Following the spatial-unit setting of Wen et al.'s MoRA: Mobility as the Backbone for Geospatial Representation Learning at Scale (ICLR 2026), we construct an alternative hierarchy analogous to the entity--block--region--city hierarchy: entity--H3 resolution 8 hexagons (block-like)--H3 resolution 6 hexagons (region-like)--city . We then project the entities and labels onto H3 resolution 6 (region-like) and resolution 8 (block-like) hexagons. Across the United States, r6 contains 273,022 hexagons with an average area of approximately 36.5 km$^2$, while r8 contains 13,217,486 hexagons with an average area of approximately 0.75 km$^2$. We freeze the representations and train linear probes under the same folds, reporting Macro-F1 for Land use and R2R^2 for the other tasks:

MethodLand use ↑Road density ↑Nightlight↑CO$_2$ ↑PM$_{2.5}$ ↑Urban heat ↑
HESTIA0.4400.7370.8560.1930.4420.296
AlphaEarth0.2880.5420.4870.0370.4400.307

HESTIA performs better on 5 of the 6 tasks, while AlphaEarth performs better on urban heat.

[Weakness 2-3: Generalization across Geographic Scopes]

We select Canada and Mexico for cross-country evaluation of HESTIA. We freeze the HESTIA checkpoint trained in the United States, perform no pretraining or fine-tuning in the target countries, run inference only on the target-country tokens, and then train lightweight probes under fixed spatial folds.

Canada uses Dissemination Areas (DAs, the census dissemination areas of Statistics Canada) as region units (57,932 units), while Mexico uses Áreas Geoestadísticas Básicas (AGEBs, Basic Geostatistical Areas) as region units (74,760 units).

Canadian regions (DAs).

MethodCO$_2$ ↑Nightlight ↑PM$_{2.5}$ ↑Urban heat ↑Population density ↑Age distribution ↓Household income ↑
HESTIA0.4300.722-1.289-0.2080.3950.2270.238
S2Vec0.3760.661-1.456-0.4520.2030.2440.128
AlphaEarth0.2530.630-0.7880.0230.2500.333-0.089

Mexican regions (AGEBs).

MethodCO$_2$ ↑Nightlight ↑PM$_{2.5}$ ↑Urban heat ↑Population density ↑Age distribution ↓
HESTIA0.1410.508-0.017-0.1950.3880.305
S2Vec0.1110.439-1.615-0.1390.0600.309
AlphaEarth0.0660.486-0.3870.589-0.5070.310

The two tables above use the same five spatial folds and valid labels, with linear probes trained on the frozen representations. HESTIA achieves the best result on 10 of the 13 cross-country task combinations (5/7 in Canada and 5/6 in Mexico), showing that the representations trained in the United States can transfer to different countries and spatial units.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Friendly Reminder for Feedback

Dear Reviewer WKK1, HdJF and QJtq,

Thanks for your attention to this message. This is a friendly reminder that the author-reviewer discussion period is ending. We would appreciate any additional feedback or suggestions you might have on our work. Your insights are valuable to us. Thank you.

Sincerely,

Authors of Submission 9628

AuthorsOpenReview ↗

Acknowledgments to the Reviewers

Dear Reviewer,

Thank you again for your constructive feedback. We have now completed and uploaded the revised manuscript, which incorporates the additional experiments, analyses, and clarifications presented during the rebuttal and discussion period.

We would be very grateful if you could briefly review the revised manuscript at your convenience. We fully understand that your time is limited and sincerely appreciate any feedback you may be able to provide.

We wish you continued success in your research and every success with your future publications.

Thank you again for your time and careful review.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.