acceptodds
Withdrawn from ICLR 2027

Puffin: Probabilistic Learning of Spatial Detail from Coarse Observations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-resolution socioeconomic variables are important for applications such as urban planning, public health, disaster response, and resource allocation. In practice, however, these variables are often observed only at a coarse spatial resolution. We introduce PUFFIN, a new paradigm for statistical disaggregation that increases the spatial resolution of coarse measurements by exploiting high-resolution satellite embeddings as auxiliary covariates. Instead of predicting a single value for each fine-resolution subregion, PUFFIN learns a probability distribution and is trained through an aggregation-aware likelihood. At inference, PUFFIN conditions these predictions on the observed regional total and splits it among the subregions. The resulting fine-scale estimates are consistent with the observed aggregate and come with calibrated uncertainty, without requiring fine-resolution labels for training. We evaluate PUFFIN on German and US census data across population, jobs, and other count variables, and study when statistical disaggregation succeeds or fails across regions, countries, and targets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
weakly supervised learning, self-supervised satellite representations, foundation model embeddings, representation transfer, spatial disaggregation, uncertainty calibration, population mapping

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QqW6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work presents Puffin, a probabilistic framework for spatial disaggregation that infers high-resolution socioeconomic estimates from coarse regional totals and satellite embeddings, without fine-grained training labels. Evaluated on German and US datasets, Puffin outperforms baselines on population and jobs. Its point estimates fail for signal-unrelated or rare events, though uncertainty intervals remain well-calibrated.

Critical strengths
  1. It investigates an interesting and practically-relevant topic of predicting fine-resolution socioeconomic metrics with only coarse aggregate statistics and satellite embeddings.
  2. The paper transparently reports failure scenarios where the Puffin framework yields poor results, offering clear practical boundaries for model application.
Critical weaknesses
  1. The paper lacks comparisons with deterministic neural allocation methods that also preserve regional totals, making it unclear whether probabilistic modeling provides additional benefits.
  2. The aggregate likelihood only constrains summed means and variances, allowing different fine-scale variance distributions to produce identical likelihoods. Thus, the reliability of learned fine-scale uncertainty remains unclear.
  3. Although training requires no fine-resolution labels, uncertainty calibration still depends on them. This raises concerns about applicability to regions without fine-resolution ground truth.
  4. The baselines are primarily traditional methods, without competitive neural disaggregation approaches. This makes it difficult to attribute the improvements specifically to the proposed framework.
  5. As demonstrated by the voter-turnout failure case in the paper, Puffin cannot produce reliable disaggregation results for outcomes such as turnout, which show weak associations with satellite-derived features.
Questions for authors
  1. Could the authors compare Puffin with a deterministic neural allocation model that also preserves regional totals, to clarify the additional benefits of the probabilistic formulation?
  2. How can Puffin provide calibrated uncertainty in regions where fine-resolution labels are entirely unavailable? Have the authors evaluated whether the calibration factor transfers across regions or countries?
Reviewer_TMh5OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

PUFFIN addresses spatial disaggregation: estimating fine-resolution counts from observed regional totals and satellite embeddings. A small neural network predicts Gaussian means and variances for individual subregions and is trained using the likelihood of regional aggregates. At inference, Gaussian conditioning distributes the discrepancy between predicted and observed totals according to the predicted variances. Clipping and rescaling enforce non-negative, mass-preserving point estimates. A two-stage procedure calibrates intervals, using coarse residuals followed by a global factor estimated from fine-resolution validation labels.

Experiments cover German population grids, US population and employment counts, and negative cases involving election turnout and gun violence. Evaluation uses spatial cross-validation, five random seeds, and a within-region R² metric. PUFFIN is competitive with linear ecological regression in Germany and substantially stronger on US population and employment, but does not improve turnout disaggregation.

Critical strengths

Critical Strengths

A simple, practical approach to learning from aggregate supervision. The framework combines a compact neural network with analytic Gaussian conditioning, avoiding fine-resolution labels for fitting the point predictor. Exact conservation of observed totals is useful for applications that require consistency with published statistics. The US results are substantial: within-region R² reaches 0.377 for population and 0.209 for jobs, compared with −1.037 and 0.050 for the ecological-regression baseline, respectively.

Broad evaluation with informative negative results. The paper evaluates multiple spatial resolutions, regular and irregular units, two countries, and several targets. The within-region metric appropriately measures the allocation of each known total rather than rewarding between-region variation already supplied by the totals. Reporting that ecological regression outperforms PUFFIN on average in Berlin, that aggregate validation can miss fine-scale overfitting, and that turnout disaggregation fails makes the empirical account more useful and credible.

Critical weaknesses

The evidence does not adequately establish the central claim of useful calibrated uncertainty. Section 3.2 clips and rescales the Gaussian conditional mean but retains variances derived from the unmodified Gaussian conditional distribution. This produces a practical estimator, but the resulting point estimate and intervals are not demonstrated to be moments or credible intervals of a coherent non-negative posterior. The phrase “posterior mean after clipping” therefore needs qualification.

The empirical validation reports approximately 90% marginal coverage after fitting a fine-label-based calibration factor. It does not report interval widths, interval scores, coverage for inhabited or high-count subregions, or comparisons with simpler calibrated uncertainty estimates. The authors themselves acknowledge that empty subregions make marginal coverage a weak test. Consequently, nominal coverage alone does not establish that the learned variances provide informative uncertainty where allocation errors matter most. Moreover, deployment with calibrated intervals where fine labels are unavailable requires evidence that the calibration factor transfers; separately evaluating countries does not establish this.

The comparisons do not isolate the benefit of probabilistic learning and variance-weighted conditioning. The baselines are uniform allocation, interpolation, and linear ecological regression. There is no matched nonlinear aggregate-supervised model with clipping and proportional rescaling, nor an equal-variance conditioning ablation. The unconditioned prior is not a sufficient control because it does not satisfy the observed aggregate constraint.

This matters because the average improvement over ecological regression is small in Germany (0.562 versus 0.544), while PUFFIN is worse in Berlin (0.432 versus 0.448; Table 1). The larger US gains could reflect nonlinear density prediction, the treatment of heterogeneous areas, variance-weighted conditioning, or their combination. The present evidence cannot distinguish these explanations. A controlled comparison is particularly important given the paper’s attribution of the German improvement to the variance-weighted split.

Questions for authors
  1. Clarification of the probabilistic interpretation: After clipping and rescaling the conditional mean, what distribution do the reported variances and intervals describe? How frequently does clipping occur, and how much does it change the estimates in each principal dataset?

  2. Evidence needed to assess uncertainty: Can you report interval widths or an interval score, together with coverage stratified by zero/nonzero counts and count magnitude? How does PUFFIN compare with an equally calibrated constant-scale or simple count-dependent uncertainty baseline? These results would determine whether the learned uncertainty adds value beyond marginal coverage.

  3. Evidence needed to assess the methodological contribution: Under identical embeddings, architecture, folds, and training supervision, how does PUFFIN compare with an aggregate-MSE neural predictor followed by clipping and proportional rescaling, and with an equal-variance conditioning variant? These comparisons would materially affect the assessment of the probabilistic formulation.

  4. Clarification of calibration and deployment: Is k fitted separately for each country, target, resolution, and cross-validation fold? Which regional residuals enter each tile’s calibration factor γₜ? Is there evidence that a fixed κ remains calibrated in a new geography without fine-resolution calibration labels?

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

PUFFIN is a spatial disaggregation method for socioeconomic variables that utilizes Earth foundation models (i.e., AlphaEarth) to learn a probabilistic distribution over unlabeled finer-scale subregion values from known coarser regional totals. A small network predicts a Gaussian prior per subregion and is trained only through the likelihood of the observed regional totals. At inference, the learned prior is conditioned on each regional total to produce a closed-form posterior that reproduces the total exactly. The posterior variances are converted into intervals through a two-stage calibration step that uses a global factor tuned to fine-scale validation labels. PUFFIN is evaluated with within-region R^2 on Germany/Berlin census, US population and jobs, and precinct turnout and gun violence (where the method is not expected to help). PUFFIN outperforms the baselines in Germany and the US, and its calibrated 90% intervals achieve marginal coverage of 0.88–0.92.

Critical strengths
  1. The problem is well motivated, and relying only on satellite embeddings that are available everywhere is a sensible choice for regions lacking fine-scale data.
  2. A probabilistic formulation is natural given the inherent uncertainty in disaggregation, and the attention to calibrated uncertainty is a strong plus.
  3. The analysis of where the method works and where it does not (e.g., precinct turnout and gun violence) is valuable.
Critical weaknesses
  1. There are a number of heuristics that are applied because of the use of the Gaussian distribution on count data (you have to address the issue of positive support for negative counts). Why not use the Poisson-Multinomial or Bayesian approach? Those seem to be the gold standard in this domain.
  2. The model also seems to ignore the possibility that there is noise in the labels, which often exists and is often well described by aggregation data products.
  3. While PUFFIN is compared against bicubic interpolation and Goodman-style regression, there are some important disaggregation baselines omitted. POPCORN [a], a satellite image-based population disaggregation is cited but not compared against. The paper also discusses gridded products that use dasymetric mapping, but do not compare against any baseline. For example, one could do a WorldPop-style approach by fitting a random forest on each region's (log) population density from its average AEF embedding, then applying it to subregional AEF and rescaling. The above methods could be implemented with the same AEF inputs and regional totals PUFFIN uses, and could better isolate the contribution of the probabilistic approach. A Poisson baseline in the spirit of [2] would help justify the choice of a Gaussian likelihood. There's also RAL [3], which is similar to the proposed prior model, although it does end-to-end optimization rather than using a pre-trained embedding.

a. Metzger et al., “High-resolution Population Maps Derived from Sentinel-1 and Sentinel-2,” Remote Sensing of Environment, 2024. b. Law et al., “Variational Learning on Aggregate Outputs with Gaussian Processes,” NeurIPS 2018.
c. Jacobs et al., "A Weakly Supervised Approach for Estimating Spatial Density Functions from High-Resolution Satellite Imagery,", SIGSPATIAL 2018.

Questions for authors
  1. What AEF embeddings are used for the tasks? Do they match the year of the aggregate data?
  2. Could the authors compare to the baselines stated in the Critical Weaknesses?
  3. How discriminative are AEF features with respect to population counts? For example, could authors show a learning-free Markov random field (MRF) baseline (used in [a])? This would assign subregions with similar AEF embeddings similar counts, while keeping each region’s total fixed.
  4. In the Berlin setting (Tables 1, 3, 7), Goodman’s regression matches or outperforms PUFFIN, even in configurations with many training regions, which overfitting may not fully explain. This feels surprising, since urban areas should offer more spatially discriminative satellite features to exploit. Could the authors use the US data to separate data quantity from urban composition, for example by comparing metropolitan tracts against random subsets of matched size?

a. Metzger et al., “Fine-grained Population Mapping from Coarse Census Counts and Open Geodata,” Scientific Reports, 2022.

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

Related papers

Discussion (0)

Sign in to comment.