acceptodds
Under review as a conference paper at ICLR 2027

A Continuous View of Urban Space: Neural Dirichlet Random Measures for Urban Spatial Learning

Abstract

Urban regions aggregate different information depending on specific partitions, but existing data-driven urban region representations are typically learned on a fixed spatial partition, limiting their applicability. To address this limitation, we propose a neural Dirichlet random measure: instead of learning representations for regions in a fixed partition, we learn a continuous intensity field over the entire spatial domain. Thus, integrating the field over any region gives its quantitative representation while ensuring conservation. Moreover, this continuous representation can also be learned directly from observations on different urban partitions, leading to a more flexible and generalizable way. We conduct experiments from both perspectives: as a predictor on arbitrary partitions, and as a learner of partition-agnostic urban spatial embeddings. As a predictor, experiments on ten major U.S. cities with widely used urban spatial embeddings show that our method guarantees quantitative conservation and also yields generally better predictive quality than conventional direct prediction. As a learner, cross-city transfer validation shows that our method learns partition-agnostic embeddings directly from observations on different partitions, and that these embeddings transfer to unseen cities and generally outperform those trained in the conventional fixed-partition way. These indicate that it is feasible to remove the dependence on fixed spatial partitions, which provides a new direction for urban representation learning.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Dirichlet random measure, probabilistic spatial modeling, geostatistics, urban foundation model, geospatial embedding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vUazOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Neural Dirichlet Random Measures (Neural DRM) for partition-agnostic urban spatial learning. Instead of learning representations for predefined spatial regions, the method constructs a continuous intensity field whose integral over arbitrary regions provides additive spatial quantities. A Dirichlet–multinomial formulation further ensures quantity conservation and probabilistic consistency across spatial partitions. Experiments on ten U.S. cities demonstrate improved population and POI prediction accuracy, as well as better cross-city transfer compared with conventional fixed-partition representation learning.

Critical strengths

S1: The paper addresses an important limitation of urban spatial representation learning: dependence on predefined spatial partitions and inconsistency across spatial scales.

S2: The integration-based formulation guarantees quantity conservation, while the Dirichlet–multinomial model provides a principled probabilistic framework for spatial count allocation.

S3: The experiments cover ten U.S. cities and multiple spatial partitions, demonstrating consistent improvements in population prediction and cross-city transfer.

S4: The paper is generally well organized, with clear mathematical formulations and extensive experimental comparisons.

Critical weaknesses
  1. Limited novelty. The method combines established techniques, including continuous intensity fields, spatial integration, and Dirichlet–multinomial distributions, without substantial methodological innovation.

  2. Insufficient baselines. Comparisons with established population disaggregation, dasymetric mapping, and spatial statistical methods are missing.

  3. Overstated partition-free claims. The implementation relies on a fixed 40 m lattice and pretrained embeddings, limiting its spatial resolution and flexibility.

  4. Limited applicability. The method requires known regional totals, restricting it to conditional count allocation rather than general urban prediction.

  5. Weak representation learning gains. Improvements for POI prediction are marginal (1%–3%), and target-city training performs better in several settings.

  6. Limited theoretical novelty. Quantity conservation follows directly from established properties of integration and Dirichlet distributions rather than new theoretical insights.

Questions for authors

See the above

6 comments
AuthorsOpenReview ↗

Response to Reviewer vUaz (1/4)

We thank you for the careful review and the valuable, constructive suggestions, and for recognising the importance of the problem and the principled formulation. You raise six points:

  1. Novelty
  2. Baselines
  3. Partition-free claims
  4. Known totals
  5. Representation-learning gains
  6. Theoretical novelty

Limited novelty: established techniques are combined without substantial methodological innovation.

We agree that the components are established: continuous intensity fields, integration over regions and the Dirichlet–multinomial distribution are all known tools.

Our contribution is to turn them into a representation-learning paradigm. These tools have been used for interpolation and allocation in single tasks. We make them a way to train urban foundation models and geospatial embeddings: one representation learns from statistics on any partition and serves any partition.

The key methodological choice is to read out at every location and then integrate over regions, instead of first pooling representations within a region and then reading out, as conventional urban representation learning does.

Controlled evidence. The paper isolates this choice: in the comparisons below, the data, the network architecture and the likelihood are the same, and only the readout differs.

  • Table 1: the additivity error drops from 0.86 to 9.3 × 10⁻¹⁰, and the MAE is lower on the training partition and on unseen partitions.
  • Table 8: with the pooled readout, pretraining on the same data gives a higher NLL than the raw input features at every label fraction, significantly in 5 of the 6 settings; with the integral readout, the encoder is the only pretrained encoder that improves on them. The gain thus comes from the integral readout rather than from the data.

What this enables. One representation predicts consistently on any partition, learns from data on any partition, with each dataset entering training on its own partition (Appendix C), and transfers to new cities (Table 2, Figure 3).

Insufficient baselines: no comparison with population disaggregation, dasymetric mapping or spatial statistical methods.

These methods allocate one quantity within one region, whereas our aim is a reusable representation that stays consistent across partitions (Problem 2). Disaggregation is still a direct test of whether the field formulation itself is sound, so we compared with them in the paper's setting and with the same protocol. The table below groups the nine methods by the three families you name; the random-forest dasymetric method is the one behind WorldPop.

MethodAlphaEarth, populationAlphaEarth, POIGHSL, populationGHSL, POI
Population disaggregationRandom-forest dasymetric (Stevens et al., 2015)+17.3%+15.6%+9.5%+13.6%
POPCORN (Metzger et al., 2024)−1.7%+9.9%−2.9%+4.0%
Dasymetric mappingAreal weighting+42.6%+18.5%+35.8%+13.3%
Ancillary-layer dasymetric (GHSL) §+10.4%+11.7%−0.4%+6.0%
Intelligent dasymetric (Mennis & Hultgren, 2006) †+29.2%+7.0%+20.6%+0.8%
Hybrid statistical-dasymetric (Sapena et al., 2022) †+13.8%+5.4%+5.2%+2.3%
Spatial statisticsPycnophylactic interpolation (Tobler, 1979) ‡+41.5%+17.3%+34.5%+11.9%
Area-to-point kriging (Kyriakidis, 2004; Liu et al., 2008) ‡+35.0%+13.6%+11.5%+8.8%
Disaggregation regression (Nandi et al., 2023)+21.2%+1.9%+9.8%+3.9%

Each entry is the reduction in MAE achieved by our field, averaged over the ten cities and the evaluation partitions; positive values favour our field. § uses GHSL layers, which are extra information in the AlphaEarth columns. † also uses NLCD land cover. ‡ also uses the totals of neighbouring tracts.

The results:

  • With AlphaEarth input, the classical methods are clearly worse than our field for population, and a few come close for POIs. With GHSL input, two of them tie with our field, as they use the same or additional information (Appendix E).
  • POPCORN has a 2–3% lower population MAE than our field; our field has a 4–10% lower POI MAE and a lower NLL in most settings.

On disaggregation, the field formulation is thus on par with the strongest method, and it also serves as a representation reused across quantities, partitions and cities.

In the revision, we have added Appendix E with the results and the details of each method, a paragraph in Section 4.2 that summarizes them, and a sentence at the end of Appendix A.2 that points to them.

AuthorsOpenReview ↗

Response to Reviewer vUaz (2/4)

Overstated partition-free claims: a fixed 40 m lattice and pretrained embeddings limit resolution and flexibility.

We agree that the wording can be misread, and we clarify what partition-free means. Neither training nor prediction is tied to a particular partition: statistics on any partition can serve as training data, and any partition can be queried. It does not mean unlimited spatial resolution.

The 40 m grid is not a partition. It only serves to compute the integrals numerically, is shared by all partitions, and can be refined. We integrated the same trained field on grids refined from 80 m to 10 m (new Appendix G). The block shares converge: for population with AlphaEarth, their total-variation distance to the 10 m result is 0.086 at 80 m, 0.037 at 40 m and 0.025 at 20 m. The conclusions are unchanged at 40 m and finer, and at 10 m the population NLL is lower than with the paper's 40 m grid in all ten cities; only the coarser 80 m grid hurts. Refining the grid only resolves region boundaries more finely; the input stays at 40 m.

Pretrained embeddings are inputs, not a limit of the method. We agree that what the field can express is bounded by the information in its input (RQ3). The input, however, is not restricted in form or source. Any modality, including images, polygons, text and graphs, enters through an encoder that gives a feature at each location, and the features of several sources are concatenated into c(x)c(x). Existing work shows that such data carry information on population and human activity:

  • images: street view (Fan et al., 2023) and satellite imagery (Jean et al., 2016; Yeh et al., 2020), including building energy consumption (Sun et al., 2026);
  • polygons: building footprints and 3D buildings (Li et al., 2023; Esch et al., 2022; Che et al., 2024);
  • text: GeoLLM (Manvi et al., 2024);
  • graphs and mobility: POI graphs and mobility flows (Huang et al., 2023; Wang & Li, 2017; Schläpfer et al., 2021; Moro et al., 2021);
  • mobile phone data (Aiken et al., 2022; Chi et al., 2022);
  • night lights (Li et al., 2020).

RQ3 already applies the same field to 12 input representations, including image-token models that take images directly (Figure 4, Table 9). We use pretrained embeddings such as AlphaEarth in the experiments because they are available worldwide, natively at 10 m, and thus support broad experiments across ten cities, several partitions and 12 input representations. Recent work keeps extending what such satellite data reveal about human activity, from mobility flows in cities (Xu et al., 2025) to human development (Sherman et al., 2026). As massive urban data are brought in, the field places no constraint on how the information is organized: new sources can be added to the input, and the field can also be learned end to end from raw data such as imagery, a direction the paper names as future work.

In the revision, we have stated this in Definition 2, in Section 3 (Lines 245–253) and in the introduction (see also our reply to Reviewer EJB1).

Limited applicability: known regional totals are required.

We agree, and the paper lists this as its first limitation. However, we want to clarify that conditioning on totals is a deliberate design. It separates how a quantity is distributed within a region from how much of it there is, so data from different sources and partitions can be learned together (Problem 2). Known totals are also common in practice, such as census counts and administrative statistics. Replacing the Dirichlet random measure with a Poisson or negative-binomial random measure would remove the need for totals; the paper names this as future work.

AuthorsOpenReview ↗

Response to Reviewer vUaz (3/4)

Weak representation-learning gains: POI improves by only 1–3%, and target-city training is better in several settings.

We agree that the POI gains in MAE are small. There are two reasons:

  1. In pretraining, POIs are observed only on block groups, about 14 blocks each, so the encoder never sees POIs at the block scale. This is a harder test than for population, which is observed on blocks.
  2. POIs are sparse: 81% of blocks have none, which compresses differences in MAE. The same holds in Table 1, where our field lowers the MAE relative to areal weighting by 36–44% for population but by 8–23% for POIs.

The NLL, which evaluates the full predictive distribution, still improves consistently for POIs: against fixed-partition pretraining, it is significantly lower at every label fraction and lower in 8 or 9 cities.

For population, the gains are large: against fixed-partition pretraining, the block MAE drops by about 15% at every label fraction, and the NLL is lower in all ten cities (Table 2, Figure 3). All these comparisons are controlled, and the gain comes from the integral readout itself. Pooled pretraining, with the same data and architecture, is significantly worse (point 1), and a random encoder with the same architecture is much worse, which rules out the extra nonlinearity as the source. Conventional pretraining is even significantly worse than the raw input features in 11 of the 12 settings, i.e., it transfers negatively (Table 8).

Target-city training uses the labels of the target city itself and serves as an upper reference. With 5% and 20% labels, our encoder has a significantly lower population NLL; with all labels, and in POI MAE, target-city training is better, as the paper reports. Pretraining thus helps most when labels are scarce.

Limited theoretical novelty: conservation follows from known properties of integration and Dirichlet distributions.

We agree that conservation follows from known properties of integration and of the Dirichlet distribution, and we do not claim it as a new theorem. Our contribution lies in using these properties for urban representation learning, so that one representation can learn from and predict on any partition (point 1).

The revision also takes the theory one step further. We extend the Dirichlet random measure from space to space × attribute, as a marked random measure (Daley & Vere-Jones, 2003), and add a proposition in Appendix F, to which Section 3.3 now points. It shows that:

  1. any indicator that is a statistical functional of additive counts, such as a median, a ratio or a Gini index, can be computed on any region from the predicted counts;
  2. these indicators are consistent across all partitions: the counts of the subregions sum to those of the region, whose indicator is thus recovered exactly;
  3. predicting such an indicator region by region cannot guarantee this, because, for example, the median of a region is not determined by the medians of its parts.

This turns conservation from a property of counts into a consistency guarantee for a much broader class of urban indicators, which covers every estimate in the ACS detailed tables (see our reply to Reviewer KdXs). It also shows what a field-based model should learn: the additive measure behind an indicator, from which the indicator follows by computation.

A preliminary demonstration. We built three such indicators from fields in the ten cities, with both inputs: median household income, median home value and per capita income. The block-group values are consistent with the tract values, and with AlphaEarth input, 73–83% of the block groups fall within the published 90% margins of error of the ACS (Appendix F). These results show feasibility only.

We present the proposition as a formal extension rather than a new theory, but it substantially widens what field-based urban foundation models can learn from. We will study this direction in more depth in future work, including indices that require modeling many variables jointly or the dependence between locations.

Reviewer_KdXsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper tends to address the limitation that exisiting urban region representation learning methods are based on fixed spatial partition, and do not guarantee the quantity conservation when the learned representations are aggregated over larger spatial units. The authors propose to learn a continuous sptail intensity field from fine-grained geospatial features (e.g., AlphaEarth embeddings) and a Dirichlet-multinomial distribution is applied to ensure quantity conservation. The framework is then evaluated as both a predictor and a learner, with the ability to generalize to unseen partitions as well as transfer to unseen cities tested.

Critical strengths
  • The paper raises an interesting and well-motivated question of whether learning continuous spatial fields could be a new direction for representation learning in urban foundation models.
  • The inclusion of negative log-likelihood (NLL) comparisons in addition to MAE is important given the probabilistic formulation of the proposed method.
  • The paper further discusses the choice of input representations for field-based urban foundation models by comparing 12 different geospatial representations, including image-based ones. However, further theoretical or empirical justification may be needed to explain why certain representations lead to substantially worse results before concluding that their features are less "informative".
Critical weaknesses
  • A key premise of this paper is that the target variable is spatially additive and should satisfy quantity conservation. However, many common urban indicators with reaslistic mearnings are not additive, such as average housing prices and median household income.
  • Although the paper emphasizes continuous field, the proposed method is actually built upon discrete but very fine-grained geospatial represenations. Such high-resolution input features may not always be available, or sometimes they are not meaningful if we are interested in human activity instead on environmental characteristics.
  • Predicting labels for finer spatial units from coarser observations (i.e., spatial disaggregation) is a well-studied problem with established methodologies. Although the paper discusses several relevant approaches in Appendix A.2, it does not experimentally compare against these methods, which limits the methodological contribution of the proposed method.
Questions for authors

Overall, I like the framing and the motivation behind the paper, which could possibly inspire a new direction for urban representation learning. My main question is that can the proposed framework be extended to non-additive urban indicators, such as average housing prices?

6 comments
AuthorsOpenReview ↗

Response to Reviewer KdXs (1/4)

We thank you for the careful review and the valuable, constructive suggestions, and we are glad that you find the framing and motivation interesting. You raise four points:

  1. Non-additive indicators
  2. Dependence on fine-grained inputs
  3. Comparison with disaggregation methods
  4. Why some representations perform worse

Extend the framework to non-additive indicators

You are right that the current paper treats additive quantities only. This is a limitation, and the revision states it in the Limitations. It also defines additive quantities in Section 2 to make this scope explicit, and points to the extension below in Section 3.3. We also agree that it is an exciting direction for follow-up work.

We would like to clarify that most urban indicators are additive counts: of the 27,990 estimates in the ACS 5-year detailed tables, 90% are additive. Many of the others are not additive themselves but are statistical functionals of additive counts, so an extension of our method covers them:

  • a mean is a ratio of two additive totals, such as aggregate income over households;
  • a median or a Gini index is a functional of the distribution of an attribute over the units of a region, such as households over income brackets.

Setting. To see why our method covers them, note that these indicators are built from individual units, such as people, households or housing units. Each unit has a location and attributes, such as income or home value, and the indicator of a region is a statistic of the attributes of the units inside it. It is therefore a known function TT of additive quantities of the region, I(A)=T(N1(A),…,NK(A))I(A) = T(N_1(A), …, N_K(A)), such as the household counts per income bracket for a median, or aggregate income and households for a mean. Our method can cover it by predicting these quantities with fields and applying TT.

Theory. We formalize this with a marked random measure MM on space × attribute, where M(A×B)M(A \times B) is the number of units in region AA whose attribute falls in BB, with one field per attribute bracket. The indicator of any region is computed from its counts. Because these counts are additive, the indicator is consistent across all partitions: the counts of the subregions sum to those of the region, whose indicator is thus recovered exactly. Predicting the indicator region by region cannot guarantee this, because the median of a region is not determined by the medians of its parts: {1,2,9}\lbrace 1, 2, 9 \rbrace and {3,4,5}\lbrace 3, 4, 5 \rbrace have medians 22 and 44, as do {2,2,2}\lbrace 2, 2, 2 \rbrace and {4,4,4}\lbrace 4, 4, 4 \rbrace, but the medians of their unions are 3.53.5 and 33. The new Appendix F states this as a proposition.

Coverage. The remaining 10% of the ACS estimates are all such functionals (quantiles, ratios and the Gini index), so the extended framework covers every ACS estimate. The full classification is in the anonymous repository.

Empirical demonstration. We built three indicators from fields in the ten cities, with both inputs: median household income, median home value and per capita income. Median home value is the ACS counterpart of housing prices and is published with margins of error; an average price is a ratio of two additive totals and is covered in the same way. The block-group values are consistent with the tract values. With AlphaEarth input, 73–83% of the block groups fall within the published 90% margins of error. These demonstrations show feasibility only: the scope of this paper remains additive quantities. Indicators that require modeling such functions are of great interest to us, but they are beyond the focus of this paper, and we will dive into them in future work. Appendix F gives the details.

AuthorsOpenReview ↗

Response to Reviewer KdXs (2/4)

Beyond ACS. Some complex indices are functionals of the same measure in principle, on subpartitions or on pairs of locations, but are hard to model in practice:

  • segregation, diversity and concentration indices (e.g., dissimilarity, entropy, land-use mix, Herfindahl–Hirschman) need joint fields of many groups over subunits;
  • urban-form indices (e.g., sprawl, polycentricity, population-weighted density) depend on both the distribution and the arrangement of population;
  • spatial autocorrelation and hot-spot indices (e.g., Moran's I, Getis–Ord) depend on the correlation between locations, and our Dirichlet structure does not capture positive correlation;
  • composite indices (e.g., the Social Vulnerability Index, the Area Deprivation Index) standardize and rank many variables against a set of regions;
  • interactions between locations (e.g., accessibility, commuting flows) are defined on pairs of locations.

They require modeling many variables jointly, the dependence between locations, and uncertainty through nonlinear operations. We leave them to future work as well. Even so, the extension lets field-based urban foundation models learn from a much wider range of statistics.

Discrete and fine-grained inputs may be unavailable or not meaningful for human activity

We agree that what the field can express is bounded by the information in its input. However, the field can draw on a very broad range of inputs, which leaves much room to raise this bound.

Inputs are flexible in form and source. Data on human activity, such as buildings, night lights, mobile phone records and mobility, can all serve as input. Any modality, including images, polygons, text and graphs, enters through an encoder that gives a feature at each location, and the features of several sources are concatenated into c(x)c(x). Existing work shows that such data carry information on population and activity:

  • images: street view (Fan et al., 2023) and satellite imagery (Jean et al., 2016; Yeh et al., 2020), including building energy consumption (Sun et al., 2026);
  • polygons: building footprints and 3D buildings (Li et al., 2023; Esch et al., 2022; Che et al., 2024);
  • text: GeoLLM (Manvi et al., 2024);
  • graphs and mobility: POI graphs and mobility flows (Huang et al., 2023; Wang & Li, 2017; Schläpfer et al., 2021; Moro et al., 2021);
  • mobile phone data (Aiken et al., 2022; Chi et al., 2022);
  • night lights (Li et al., 2020).

RQ3 already applies the same field to 12 input representations, including image-token models that take images directly (Figure 4, Table 9).

Because such data are not always available, we start from globally available remote sensing, e.g., AlphaEarth at 10 m every year and GHSL. Remote sensing captures human activity, especially through the built environment (Jean et al., 2016; Yeh et al., 2020; Rolf et al., 2021; Burke et al., 2021). Our results agree: with AlphaEarth or GHSL input, our field lowers the population MAE by 36–44% compared with areal weighting (Table 1).

The gap and recent progress. Remote sensing observes the physical environment, and its relation to income or travel is indirect. Recent work is narrowing this gap. It predicts fine-grained mobility flows in cities from satellite imagery (Xu et al., 2025), estimates the Human Development Index of municipalities worldwide from provincial statistics and satellite imagery (Sherman et al., 2026), and estimates wealth and poverty by combining satellite, phone and social-network data (Chi et al., 2022; Smythe & Blumenstock, 2022).

Our method helps close the gap. Our encoder is pretrained on counts of human activity, population and POIs. In new cities, it is the only pretrained encoder that improves on the raw AlphaEarth features, while conventional pretraining falls below them (Table 8). Geospatial embeddings provide spatial detail, statistics on human activity give it meaning, and the field connects the two.

AuthorsOpenReview ↗

Response to Reviewer KdXs (3/4)

Methods of spatial disaggregation introduced in Appendix A.2 are not compared experimentally

These methods allocate one quantity within one region, whereas our aim is a reusable representation that stays consistent across partitions (Problem 2). Disaggregation is still a direct test of whether the field formulation itself is sound, so we compare with them on it.

We compared our field with nine methods from three families in the paper's setting and with the same protocol. The methods cover those in Appendix A.2:

  • implemented: Tobler (1979), Mennis & Hultgren (2006), Stevens et al. (2015) and POPCORN (Metzger et al., 2024), a follow-up of Metzger et al. (2022), in the table, and the aggregated Gaussian process of Law et al. (2018), trained with our objective (see the results);
  • the model of Tanaka et al. (2019) belongs to the same aggregated-GP family as Law et al. (2018);
  • Okawa et al. (2019) need the locations of individual events, which our setting does not observe.
MethodAlphaEarth, populationAlphaEarth, POIGHSL, populationGHSL, POI
Population disaggregationRandom-forest dasymetric (Stevens et al., 2015)+17.3%+15.6%+9.5%+13.6%
POPCORN (Metzger et al., 2024)−1.7%+9.9%−2.9%+4.0%
Dasymetric mappingAreal weighting+42.6%+18.5%+35.8%+13.3%
Ancillary-layer dasymetric (GHSL) §+10.4%+11.7%−0.4%+6.0%
Intelligent dasymetric (Mennis & Hultgren, 2006) †+29.2%+7.0%+20.6%+0.8%
Hybrid statistical-dasymetric (Sapena et al., 2022) †+13.8%+5.4%+5.2%+2.3%
Spatial statisticsPycnophylactic interpolation (Tobler, 1979) ‡+41.5%+17.3%+34.5%+11.9%
Area-to-point kriging (Kyriakidis, 2004; Liu et al., 2008) ‡+35.0%+13.6%+11.5%+8.8%
Disaggregation regression (Nandi et al., 2023)+21.2%+1.9%+9.8%+3.9%

Each entry is the reduction in MAE achieved by our field, averaged over the ten cities and the evaluation partitions; positive values favour our field. § uses GHSL layers, which are extra information in the AlphaEarth columns. † also uses NLCD land cover. ‡ also uses the totals of neighbouring tracts.

The results:

  • With AlphaEarth input, the classical methods are clearly worse than our field for population, and a few come close for POIs. With GHSL input, two of them tie with our field, as they use the same or additional information (Appendix E).
  • POPCORN, the strongest modern method, has a 2–3% lower population MAE than our field; our field has a 4–10% lower POI MAE and a lower NLL in most settings.
  • As two controlled variants, we trained POPCORN's architecture and the aggregated GP with our objective. Both are on par with our field (within ±2.7% in MAE), so the gain comes from the field formulation rather than from a particular network.

On disaggregation, the field formulation is thus on par with the strongest method. It also serves as a representation that is reused across quantities, partitions and cities, which these methods were not designed for.

In the revision, we have added Appendix E with the results and the details of each method, a paragraph in Section 4.2 that summarizes them, and a sentence at the end of Appendix A.2 that points to them.

Reviewer_EJB1OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a continuous representation of discrete spatial occurrences, aiming at mitigating the representation dependence on specific spatial partitions.

Critical strengths

None.

Critical weaknesses

W1. The authors state clearly in Line 224-230 that "In practice, we evaluate the integral on a fixed lattice of atoms {xj} with areas {wj}, shared by all regions". This reduces the entire proposed Dirichlet representation to a per-cell weighted one-hot vector representation -- the cell is one piece of their lattice. What is the difference of turning the previous partition into a raster image? From my research experience, if one tends to represent discrete objects as continuous distributions, they need to relax the boundaries (e.g. using very tight Gaussian distributions to represent Dirac functions) and make sure they are parameterized, so that the distributions can be explicitly computed. Such transformations make meaningful approximations. The method proposed in this paper is not useful in that sense.

Questions for authors

NA

8 comments
AuthorsOpenReview ↗

Response to Reviewer EJB1 (1/4)

We thank you for the careful review and the valuable, constructive comments. You raise concerns including the following four parts:

  1. Evaluating the integral on a fixed lattice
  2. How this differs from turning the partition into a raster image
  3. Representing discrete objects continuously
  4. Whether the method is useful

Evaluating the integral on a fixed lattice (Lines 224–230) reduces the representation to a per-cell weighted one-hot vector.

The representation is the intensity field a(x)=f(c(x))a(x) = f(c(x)): one neural network ff applied to the context c(x)c(x) of each location. It can be evaluated wherever the context is defined. The lattice of Lines 224–230 is only a grid on which the field is evaluated to compute the integral ∫Aa(x)dx\int_{A} a(x) dx numerically.

Main Difference: On the grid, a region is indeed a weighted indicator vector over cells, but this vector only describes the region; the values summed over it come from the field. A one-hot vector carries only the identity of a cell, so any model built on it must learn a separate value for each cell. Our field is instead a function of location: the same network maps the context c(x) at every location x to a value, which changes with c(x) and, as the network is continuous, varies continuously wherever c(x) does. The quantity of a region is the integral of this function over the region. Finer grids converge to the same integral (consistency). Because the network is shared, the field transfers to locations and cities never seen in training (RQ2) and can be integrated over any partition with exact consistency.

One limitation of the implementation remains. Our experiments aggregate the context to 40 m, so the input carries no information on variation within a 40 m pixel. The limit thus lies in the information of the input, not in the representation, which aims to make training of urban foundation models or geospatial embeddings more flexible. With a finer input, such as AlphaEarth at its native 10 m, the same field can vary more finely without any change to the method.

We agree that the paper did not make this distinction clear. In the revision, we have:

  • stated in Definition 2 that the context may come from any source of spatial information, whose resolution is finite, so the field, although defined everywhere in continuous space, can only express the spatial variation that its input carries;
  • stated in Section 3 (Lines 245–253) that the grid only serves to compute the integral numerically and is not part of the representation, and that the integrals converge as the grid is refined (new Appendix G);
  • stated in the introduction that the field is a shared function of local context that can be evaluated anywhere, and that region integrals are computed numerically;
  • made explicit in the introduction how the method follows from the motivation: the field is built for flexible urban foundation models and geospatial embeddings that learn from statistics on any partition.
AuthorsOpenReview ↗

Response to Reviewer EJB1 (2/4)

How does this differ from turning the partition into a raster image?

The difference lies in where each cell's value comes from:

  • A raster maps each cell to its own value. The model is the grid itself. Rasterizing a partition only spreads each region's value over its cells and adds no information about variation within the region.
  • Our field is a function over continuous space. It maps every location x to a value through its local context, a(x) = f(c(x)), with one network shared by all locations. It does not depend on any grid; the grid only serves to numerically compute the integral.

This leads to three differences:

  1. Generalization. A raster has no value for a cell it has never observed, and its values cannot be carried to another city. The field computes the value at any location from its context. It therefore predicts in test tracts held out by spatial folds (Table 1), and it transfers to new cities: against fixed-partition pretraining, it lowers the block MAE by about 15% for population at every label fraction, with a lower NLL in all ten cities (Table 2, Figure 3). It also transfers to unseen partitions without retraining (Figure 3).
  2. Learning from totals. With a separate value for each cell, region totals cannot determine a raster; a shared function can be learned from them across many regions.
  3. Resolution independence. Refining the grid converges to the same integral. A raster is the model itself and has no such limit.

Beyond the theoretical analysis above, we empirically compared our field with this raster. Predicting from the rasterized partition, which spreads each region's total uniformly over its cells, is areal weighting (Goodchild & Lam, 1980), the "By areasize" baseline in Table 1. Our field lowers the MAE against it by 36–44% for population and 8–23% for POIs. In both MAE and NLL, it is significantly better in every city on blocks, and in 7 to 10 of the 10 cities on unseen partitions. Rasters weighted by ancillary layers (dasymetric mapping) are also worse (Appendix E).

AuthorsOpenReview ↗

Response to Reviewer EJB1 (3/4)

Representing discrete objects continuously requires relaxed, parameterized distributions, e.g. tight Gaussians for Dirac functions.

We thank you for this point and agree that the paper did not state clearly what is observed and what is continuous.

Our data are counts on regions; the locations of individual objects are not observed. The population is available only as block counts. POIs do have exact locations (Lines 350–351 in the revision), but the model never sees them: they only serve to count POIs in regions, the blocks for training and the grid pieces for evaluation. There are therefore no Dirac functions to relax. This is the norm in cities: statistics on people, households, and jobs are published only as counts on areal units, such as census blocks, block groups or districts, because individual locations are confidential (e.g., Title 13 of the U.S. Code for the census). Exact point locations are thus unavailable for most urban quantities, and our method is designed for this setting.

Our continuous object, the intensity field, already meets the requirements you mention: it is parameterized by a neural network and can be computed explicitly. The only approximation is the numerical integral on the grid, and it is a meaningful one, since it converges as the grid is refined. We integrated the same trained field on grids refined from 80 m to 10 m:

  • For population with AlphaEarth, the block shares converge: their total-variation distance to the 10 m result is 0.086 at 80 m, 0.037 at 40 m and 0.025 at 20 m.
  • The conclusions are unchanged at 40 m and finer; only the coarser 80 m grid hurts.
  • At 10 m, the population NLL is lower than with the paper's 40 m grid in all ten cities.

To make the difference concrete, the table below compares the relaxation you describe with our method, from the information each starts from to the output it gives.

Relaxing discrete objects (e.g., tight Gaussians)Neural DRM (ours)
Known informationLocation of every objectCounts on the regions of one or more partitions; individual locations unknown
Continuous objectA kernel around each known pointThe intensity a(x)=f(c(x))a(x) = f(c(x)) over space
ParametersOne kernel per objectOne function f shared by all locations
DiscretizationNot needed for known pointsGrid only for computing ∫Aa(x)dx\int_A a(x) dx numerically; refining it converges
GoalA smooth representation of given objectsLearn from counts on any partition; predict on any partition
OutputDensity of the given objectsCount distributions consistent across partitions

The two approaches start from different information. Kernel relaxation assumes that every object's location is known and smooths each Dirac function into a density. In our setting, these locations are not observed, and only counts on regions are. The continuous object is therefore the intensity of the quantity: one function of local context, learned from the counts and integrated over any region. The grid only serves to compute this integral. In the revision, the problem definition states explicitly that only counts on regions are observed and that individual locations are unknown.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comments

We thank the reviewers and the Area Chair for their time and the constructive discussion. The reviewers' suggestions have helped us improve this work considerably. We summarize the strengths that the reviewers identified, the improvements made following their suggestions, and the outlook.

Strengths identified by the reviewers.

  • The paper addresses an important limitation of urban representation learning, the dependence on predefined partitions and the inconsistency across scales (vUaz), and its framing "could possibly inspire a new direction for urban representation learning" (KdXs).
  • The integral formulation guarantees quantity conservation, and the Dirichlet–multinomial model gives a principled probabilistic framework (vUaz).
  • The experiments cover ten cities and multiple partitions, report NLL in addition to MAE, and compare 12 input representations (vUaz, KdXs).
  • With the revision, "it is easier to understand what the proposed method is addressing" (EJB1).

Improvements following the reviewers' suggestions.

  • A comparison with nine methods of population disaggregation, dasymetric mapping and spatial statistics (KdXs, vUaz; Appendix E).
  • An extension to non-additive indicators, with a proposition, the coverage of the ACS tables and a demonstration on three indicators (KdXs, vUaz; Appendix F).
  • A clearer statement that the lattice only computes the integral numerically, with a convergence study (EJB1, vUaz; Appendix G), and that only counts on regions are observed (EJB1).
  • Two measurable diagnostics that explain why input representations differ (KdXs; Section 4.4).

Outlook. Urban foundation models and foundational geospatial embeddings, such as AlphaEarth, are meant to be reused across tasks, places and scales. Yet learning from urban statistics has required fixing a partition first, and predictions on different partitions need not add up. Our field learns from statistics on any partition and conserves quantities exactly across scales: the prediction for a region always equals the sum over its subregions. One representation can thus be reused across quantities, partitions and cities, which offers a flexible way to scale such models with larger networks and more data sources. Extending the experiments beyond population and POIs, as Reviewer EJB1 points out, is the next step in this direction; Appendix F takes a first step with census income and housing data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.