acceptodds
Under review as a conference paper at ICLR 2027

MIND the Gap: A Geographic Implicit Neural Representation With Adjustable Spatial Scale

Abstract

Geographic measurements are often sparse, leaving large areas without labels for the quantities we want to map. Geographic implicit neural representations (INRs) provide coordinate-based embeddings that can be combined with sparse labels to predict at unsampled locations without satellite imagery at inference. Yet existing INRs are largely evaluated with random holdouts, leaving their ability to generalize across larger geographic gaps unclear. We introduce Matryoshka Implicit Neural Distillation (MIND), a geographic INR whose spatial granularity can be adjusted after training. MIND distills several pretrained geospatial models using nested supervision at increasing embedding dimensions, dividing the representation into contiguous chunks. Early chunks capture broad spatial patterns, while later chunks add increasingly local variation. Downstream models can retain only the leading chunks or use our Chunked Penalty to reduce reliance on later chunks without retraining the INR. We also introduce CoordBench, comprising datasets and targets with both random and regional holdouts at multiple spatial scales. Across CoordBench, fine-scale features help most when labels are nearby, while smoother representations generalize better across larger geographic gaps. MIND with the Chunked Penalty achieves the highest aggregate regression and classification performance among tested INRs and the highest overall performance under regional holdout. These results show that geographic representations should be evaluated and adapted according to the spatial separation between labeled and prediction locations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geospatial representation learning, Implicit neural representations, Spatial generalization, Remote sensing, Knowledge distillation, Matryoshka representation learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_KSK9OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MIND (Matryoskha Implicit Neural Distillation), an earth embedding method with adjustable spatial scale. To do this, the model distills 4 pretrained geospatial embedding teachers (covering factors like climate, biodiversity or satellite imagery) through nested supervision over embedding prefixes. Coarser dimensions capture broader spatial patterns while later dimensions add finer variation. At inference time, downstream predictors can choose a preselection of dimensions or use the proposed chunked penalty to reduce reliance on the finer-scale chunks.

The paper is further accompanied by a benchmarked called CoordBench, composed of dozens of datasets and targets, evaluated with cross validation across random and regional heldouts. In this benchmark, MIND+CP achieves pareto-optimal performance in both regression and classification; while other experiments in the paper provide evidence that different tasks may benefit differently from different levels of detail in the representation.

Critical strengths
  • This paper tackles a key problem in a growing field of Earth AI, which is the level of detail in geospatial embeddings. I find both the topic and the approach to be highly timely and relevant.
  • The method is useful, well-designed, and practical. MIND provides an easy and cost effective way to adjust granularity after pretraiining, which is supported by truncation, variograms, controls and regularisation experiments.
  • THe Chunked penalty improves regional prediction while preserving performance on random folds.
  • The proposed benchmark is also a highlight in this paper, being very comprehensive in both datasets and tasks, and highly robust in how heldout performance is measured. This benchmark can be a highly-used resource for the future.
Critical weaknesses
  • I found it hard to exactly point to what is making the model work. MIND distills 4 geospatial embeddings, but the AEF contribution is difficult to measure as it is excluded from the concatenation baselines. It would be nice to strenghen the experimental setup on this regard, to fully isolate and understand the influence of each teacher under every experiment.
  • The proposed benchmark may be biased towards the MIND embedding. Particularly, many tasks are centered around cities, which could match the MIND's training coordinates. Other INR may not have this bias. This should be studied further and be clarified by the authors.
  • Most INR use 256-512 dimensional embeddings, while MIND goes way beyond this. This makes it difficult to compare exactly with other lower-dimensional embeddings. The fisher information (eg Measuring the intrinsic dimension of Earth representations, ICLR 2026) could be reported for better analyses. This warrants further analysis and discussions.
Questions for authors
  • How many MIND encoders were trained, and what is the variability across encoder seeds?
  • How much of the improvement depends on AEF?
  • The discussion could consider TerraNova: A Foundation Model for the Anthropocene, which aligns a shared representation with a set of pretrained geospatial embeddings alongside environmental and societal supervision.
2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for the careful review and for pointing us to TerraNova and the intrinsic-dimension analysis. We have added both, along with the teacher and sampling comparisons you requested.

Q1: How many MIND encoders were trained, and what is the variability across encoder seeds?

We used one released encoder for all main results which can be found at the anonymous Hugging Face link in the Reproducibility Statement. During evals we vary the folds across 5 seeds while keeping the encoder frozen. To check pretraining variability we reran pretraining with two more random seeds. For MIND+CP the variation across the three runs is smaller than across the five fold assignments at every regional block size. Notably, we found that our released encoder is also the weakest of the three runs with the two new runs producing even stronger results for MIND.

Q2: How much of the improvement depends on AEF?

We retrained MIND without AEF and random-fold R2R^2 changes by 0.001 while random-fold land-cover accuracy is nearly unchanged. We evaluate across 5 seeds and accuracy changes by at most 0.6 points and R2R^2 by at most 0.034 across the block sizes. This suggests that AEF itself isn't the main source of MIND's performance.

Q3: The discussion could consider TerraNova: A Foundation Model for the Anthropocene …

Thanks for pointing us to this paper. We added both TerraNova spatial and spatiotemporal encoders as baselines to Table 3 and cite it in Sec. 2. Its supervision includes variables that match or proxy 36 of CoordBench's 78 targets (see Table 20). MIND+CP and MIND-256 score above both TerraNova encoders on both metrics (e.g. 0.184 vs. 0.028 R2R^2 at 40° for MIND+CP).

W1: I found it hard to exactly point to what is making the model work. … the AEF contribution is difficult to measure as it is excluded from the concatenation baselines. …

We added this comparison in App. G. On regression, AEF scores below MIND+CP at every block size (0.287 vs. 0.356 R2R^2 at 10°), while AEF and TESSERA score higher on 13 of the land-cover datasets directly from AEF's evaluation suite.

Table 12 compares MIND+CP with each teacher alone, with the teachers concatenated (with and without AEF), and with a PCA of the four teachers. The concatenation that includes AEF has higher land-cover accuracy and higher R2R^2 under random folds and 2° blocks. MIND+CP has higher R2R^2 when the held-out regions are 10° or larger (0.356 vs. 0.291 at 10°).

We also retrained MIND leaving out one teacher at a time (see Q2). Without Climplicit, GeoCLIP, or SINR, random-fold R2R^2 drops by 0.029–0.038. Without AEF, random-fold R2R^2 changes by 0.001 and land-cover accuracy by at most 0.6 points. The Chunked Penalty improves regional R2R^2 for every leave-one-out teacher set at all block sizes.

W2: The proposed benchmark may be biased towards the MIND embedding. …

This is a fair concern. CoordBench is not sampled from MINDSET, and many of its datasets are global grids or station networks (e.g. WorldClim, SoilGrids, SatCLIP, air temperature). In Table 17 we now list each dataset's coverage and spatial units. To test the effect of city-centered sampling, we added an evaluation on rural points, more than 50 km from any city of at least 15,000 people (App. J, Table 14). MIND+CP scores highest on these rural points compared to SINR, GeoCLIP, and Climplicit (0.895 R2R^2 and 94.4% accuracy), even though MINDSET is sampled around cities.

W3: Most INR use 256-512 dimensional embeddings, while MIND goes way beyond this. … The fisher information (eg Measuring the intrinsic dimension of Earth representations, ICLR 2026) could be reported …

Truncated to 128 dimensions, MIND scores 0.327 R2R^2 at 10°, above every baseline at its full width; the best is GeoCLIP at 512 dimensions with 0.303. Under random folds, MIND-256 scores 0.556 R2R^2, above all six 256-dimensional baselines (at most 0.546 for TerraNova-S; Table 3).

We followed the approach of Rao et al. (ICLR 2026) and report 5 intrinsic-dimension estimators and the participation ratio in App. K. We found the largest difference between encoders is in the participation ratio which counts the linear directions an embedding uses. At the same width, MIND’s is the highest, with 64 at 256 dimensions vs. at most 43 for the six baselines of the same width, 123 vs. 63 for GeoCLIP at 512, and 244 vs. 89 for Climplicit at 1024. The participation ratio is also correlated with higher CoordBench scores (Spearman 0.78–0.84). In contrast, we found FisherS is about 10 or less for every encoder regardless of its embedding dimension.

Reviewer_KSK9OpenReview ↗

Response to rebuttal

I thank the authors for the detailed response to my review and for the rebuttal to the other reviews, which mentioned limitations that I missed, like missing analyses of related work. These rebuttals have substantially addressed my original concerns and increased my confidence about recommendign acceptance.

The analyses of the teacher ablations, the comparisons with concatenated teachers, smaller embedding dimensions, Terranova, and the rural and non-US evaluations are all signfiicant, very informative, and relevant, and suggest that MIND is likely the SOTA geospatial embedding across several dimensions.

Some limitations still remain, including variation across task categories, but the authors are candid about these points. I would also suggest to strenghen the analysis of the embedding distillation in MIND compared to what was done in Terranova, and where the differences lie, beyond the actual embeddign choices that were used for distillation.

I find the paper to be very interesting and useful and the evaluation to be unusually thorough and complete, and I'll be happy to keep or increase my rating. I'm also interested in knowing the thoughts of the other reviewers.

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

“Please MIND the gap between labelled training locations and test locations.” I like the title.

This paper addresses an important but often overlooked problem in GeoAI: whether geographic INRs can generalise when test locations are geographically distant from labelled training locations. Due to spatial autocorrelation, random train-test splits often place test points close to training points, primarily evaluating local interpolation rather than cross-regional generalisation. The authors propose MIND, which distils four pretrained geospatial models into a 3,072 dimensional coordinate embedding. Through nested supervision and a CP, the method aims to adapt the spatial granularity of the representation without retraining the encoder. I’m torn between 2 and 3.

Critical strengths
  1. The paper highlights a significant limitation of random train-test splits in geospatial learning. Evaluating geographic generalisation across different spatial separations is both necessary and valuable.

  2. The combination of nested distillation and scale-aware regularisation is interesting. The extensive evaluation across multiple datasets and spatial scales, together with the ablation studies, provides useful empirical evidence.

Critical weaknesses
  1. Although the paper compares several geographic INRs, MIND benefits from four heterogeneous teachers, whereas most baselines use narrower sources of pretraining information. The teacher-concatenation baseline also excludes AEF, making it difficult to isolate the benefits of the proposed method. Comparisons with classical spatial methods and explicitly multiscale coordinate encodings would further strengthen the evaluation.

  2. The main components, including Matryoshka Representation Learning and multi-teacher distillation, are well established. The proposed Chunked Penalty is essentially a weighted ridge regularisation, mathematically equivalent to feature rescaling followed by standard ridge regression. The more interesting contribution lies in the coarse-to-fine spatial ordering induced by nested distillation, rather than the penalty formulation itself.

  3. Although MIND+CP achieves strong aggregate performance, the improvements vary considerably across task categories. Furthermore, regional holdout does not necessarily guarantee strict spatial extrapolation, particularly as the encoders are pretrained on geographically distributed data.

Questions for authors
  1. How does MIND compare with other methods that explicitly control spatial scale, such as frequency-truncated Fourier features, spherical harmonics, or frequency-aware regularisation?

  2. How much of the improvement comes from combining multiple pretrained teachers, and how much from the proposed nested representation and spatial-scale regularisation? Additional teacher ablations would help clarify this.

  3. The results in Supplementary Table 7 show considerable variation across task categories. For example, under the 40° regional holdout, MIND+CP achieves an R² of 0.003 on socioeconomic tasks, compared with 0.038 for GeoCLIP. Could the authors explain this difference?

  4. Regional holdout does not necessarily represent strict spatial extrapolation. Although the buffered-exclusion experiment is helpful, could the authors further clarify the extent to which the results demonstrate generalisation to geographically unseen regions, rather than simply to locations distant from downstream training labels?

  5. Why does the non-nested full model outperform MIND Full (without CP) under the 10° and 40° regional holdouts? Could the authors explain why nested supervision appears to weaken the full representation's regional generalisation while making the Chunked Penalty more effective?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for the thoughtful review. We answer Q3 and Q4 under W3.

Q1: How does MIND compare with other methods that explicitly control spatial scale …?

Our results already include three models with explicit multiscale positional encodings: SatCLIP uses spherical harmonics up to degree 40, GeoCLIP uses random Fourier features at three scales, and CSP uses Space2Vec grid-cell sinusoids. MIND+CP scores above all three at every block size for both regression and classification datasets. Applying the Chunked Penalty to GeoCLIP gives no gain (Table 4). We think this shows that a multiscale input encoding is not enough and the penalty's benefit comes from the coarse-to-fine ordering of MIND’s dimensions produced by nested training.

To test frequency control directly, we added two fixed encodings (App. I, Table 13): spherical harmonics ordered by degree and multiscale random Fourier features ordered by length scale. For spherical harmonics the penalty increases R2R^2 from 10° on (0.108 to 0.206 at 10°, −0.298 to −0.107 at 40°) but it doesn't increase for Fourier features. This indicates the Chunked Penalty can help a fixed encoding ordered from coarse to fine. However MIND+CP still scores above both methods across the board.

Q2: How much of the improvement comes from combining multiple pretrained teachers …?

In Table 4, we compare MIND and a non-nested version with the same teachers, data, and architecture. Nesting with the Chunked Penalty raises R2R^2 over the non-nested control from 0.338 to 0.356 at 10° and from 0.127 to 0.184 at 40°. We also retrained MIND without each teacher in turn and compared each run with a four-teacher control trained with identical settings. Removing Climplicit, GeoCLIP, or SINR lowers random-fold R2R^2 by 0.029–0.038, while removing AEF changes it by 0.001. Across all block sizes, removing AEF changes R2R^2 by at most 0.034 and land-cover accuracy by at most 0.6 points. The Chunked Penalty improves regional R2R^2 for every teacher set at every block size.

Q5: Why does the non-nested full model outperform MIND Full (without CP) …?

We think nesting concentrates the broad scale signal in the leading chunks. The characteristic distance falls from 790 km for the first chunk to 200 km for the full embedding (Sec. 5.2), and MIND-128 scores 0.327 at 10° compared with 0.280 for MIND-Full. After standardizing each dimension, standard ridge penalizes all dimensions equally and relies (possibly overfits) on finer-scale features that transfer poorly to distant regions. In the non-nested model the signal is spread across the dimensions so probing without the Chunked Penalty loses less as this model's characteristic distance is about 230 km (App. D).

W1: … The teacher-concatenation baseline also excludes AEF … classical spatial methods and explicitly multiscale coordinate encodings …

Q2 gives the nested vs. non-nested control. We also evaluated AEF on CoordBench and it scores 0.287 R2R^2 at 10° and 0.047 at 40°, compared with 0.356 and 0.184 for MIND+CP (App. G). MIND+CP has higher R2R^2 than each individual teacher and the concatenation of SINR/GeoCLIP/Climplicit teachers (Table 12). Adding AEF to the concatenation gives higher accuracy and higher R2R^2 under random folds and 2° blocks, but MIND+CP has higher R2R^2 from 10° on (0.356 vs. 0.291 at 10°).

For classical spatial methods, we added ordinary and indicator kriging, which show similar results to Coord. IDW with a high random-fold score, but this gain disappears as block size increases (Table 3, App. L). The two explicit multiscale encodings are covered in Q1.

W2: … Chunked Penalty is essentially a weighted ridge regularisation … The more interesting contribution lies in the coarse-to-fine spatial ordering …

We agree on both points and now state how it's equivalent in Sec. 3.2. We agree that the coarse-to-fine ordering is the primary contribution and, whether using feature scaling or a penalty, the goal is to take advantage of the ordering not present in other INRs. Among the pretrained INRs we tested, the penalty benefits only MIND and provides no gains for GeoCLIP, the non-nested model, or MIND with randomly permuted dimensions (Table 4).

W3: … improvements vary considerably across task categories … regional holdout does not necessarily guarantee strict spatial extrapolation …

We agree. Sec. 5.1 now notes that GeoCLIP has higher socioeconomic R2R^2 at 20° and 40° and we describe MIND+CP's advantage as an aggregate result. We think that because MIND has to reconstruct all four teachers it keeps a compromise among them and does not retain all of a teacher's domain expertise, e.g. GeoCLIP's regional socioeconomic signal.

We have also clarified in Sec. 4 that the splits withhold downstream labels, while the encoders are pretrained on globally distributed data. The results therefore concern regions without downstream labels, not regions unseen in pretraining.

Reviewer_ugUGOpenReview ↗

Weak acceptance

Thank you to the authors for their response and the revised version. I am maintaining my score.

Reviewer_5pySOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MIND is a geographic neural representation trained by distilling four pretrained embeddings: AlphaEarth Foundations (satellite imagery), Climplicit (climate), GeoCLIP (geotagged photographs) and SINR (species observations). Training uses MINDSET, a new set of 12M land coordinates concentrated around cities. Separate linear heads reconstruct every teacher from the leading 64, 128, 256, etc. dimensions. The authors observe an emergent property: the leading chunks vary smoothly in space, while later chunks add progressively more local variation (about 200 km for the full embedding). They exploit this ordering at probing time.

A second contribution is the evaluation benchmark (CoordBench) based on 52 public datasets (socioeconomic, environmental and land cover), curated into a common format. The main empirical findings are: (i) Under random folds, 93% of test points lie within 25 km of a training point, and inverse-distance weighting on raw coordinates beats every tested implicit neural representation (INR). (ii) Under regional holdout, MIND+CP is the best INR at every block size in aggregate. (iii) CP helps only embeddings whose dimension ordering comes from nested training.

Critical strengths

A well-documented benchmark. The random-fold vs. regional-holdout analysis (Fig. 5, Table 3) shows that current INR benchmarks mostly measure local interpolation: a coordinate IDW baseline beats every learned encoder under random folds. The buffered-exclusion experiment (App. E, Table 5) separates the effect of train–test distance from the effect of which regions are held out. This is known in ecology and geostatistics, but it is perhaps under-appreciated in the location-encoder literature, and thus the paper does a service to this community.

A clean ablation of the role of dimension ordering. The non-nested model, the permuted-dimension control and CP applied to GeoCLIP (Table 4) together make a convincing case that CP's benefit on regression comes from the ordering induced by nested training.

Noticing that nested supervision produces a coarse-to-fine spatial decomposition (Figs. 6, 7) despite not being told about spatial scale. Being able to choose the spatial scale of one frozen representation at probing time is practical.

The comparison covers eleven pretrained INR families plus non-parametric baselines. Releasing code, the benchmark, the pretraining set, the checkpoints and a global 1 km embedding grid will be useful to others.

Critical weaknesses

Missing PCA baseline. With linear heads and an MSE/cosine loss, the first m dimensions are pushed to reconstruct the teachers as well as possible, favoring their dominant principal subspace, plausibly the spatially smooth part. The spectral-bias explanation (l. 399–401) is not tested. PCA of the concatenated teacher embeddings (and of each baseline), truncated by explained variance, is the control I would want to see. If it matches MIND+CP, nested distillation reduces to a compact coordinate-only approximation of the teachers: useful, but a different claim.

Generality vs. evaluation scope. MIND is called "a general-purpose coordinate representation" (l. 53), yet CoordBench has no species-distribution targets, although SINR is a teacher and public benchmarks exist (eBird Status & Trends, IUCN ranges, GeoLifeCLEF). MINDSET samples only around cities of at least 15,000 inhabitants (App. C), leaving rural areas largely unsampled. Either evaluate these regimes or narrow the claim.

Incomplete comparison with the teachers. The teacher-concatenation baseline (App. A, Table 8) beats MIND_Full at 2° (0.496 vs. 0.492) but was not given truncation or CP. AlphaEarth is excluded because querying it "requires access to the raster"; yet, if I am not mistaken, AEF embeddings are public and MINDSET is itself built by sampling AEF pixels. Direct teacher lookups, including AEF, with the same scale controls, are the obvious reference for what distillation adds.

The regional advantage comes from environmental targets, where overlap with the teachers is most likely. At 40° (Table 7), MIND+CP leads on the 10 environmental datasets (0.638 vs. 0.499 for Climplicit), which include WorldClim, air temperature, elevation and soil, signals close to what Climplicit and AEF encode. On the 26 socioeconomic datasets it is below its own teacher GeoCLIP at 20° (0.165 vs. 0.203) and 40° (0.003 vs. 0.038). "Best at each spatial scale" thus holds only in aggregate, and the abstract should say so. Also, all encoders are pretrained globally and only probe labels are held out (l. 315), so "generalize across larger geographic gaps" overstates the result.

Modest benefit of nesting, none for classification. At 10°, a non-nested model with plain ridge reaches 0.338 vs. 0.356 for MIND+CP, while MIND_Full without CP (0.280) is worse than the non-nested model: CP largely offsets a cost nesting introduces. For classification nesting gives no gain (0.600 vs. 0.599 at 10°). These comparisons belong in the main text, with paired tests.

Mean R^2 weights datasets equally, so 12 correlated CDC PLACES measures count as 12 datasets while PDFM's 27 targets count as one. Per-category results and win rates against the strongest baseline belong in the main text, I would say. Land-cover classification should report balanced accuracy or macro-F1.

Weak spatial baselines. IDW is the only one, a low bar. Relevant tools from ecology: nearest-neighbour-distance-matched CV (Milà et al., 2022), area of applicability (Meyer & Pebesma, 2021), and design-based validation as an alternative to pessimistically biased block CV (Wadoux et al., 2021).

Novelty of the evaluation. The claim that INRs are "largely evaluated with random holdouts" (l. 15, 44–45) overlooks existing spatially blocked linear-probe protocols, e.g., SINR's "Geo Feature" task. CoordBench's real addition is many targets and multiple scales.

Questions for authors
  1. How does a PCA-ordered baseline compare with MIND+CP and with truncated MIND?
  2. On socioeconomic targets, MIND+CP falls below GeoCLIP at 20° and 40° (Table 7). Why does distillation lose the regional signal of this teacher?
  3. Could you evaluate MIND on species-distribution tasks, e.g., eBird Status & Trends / IUCN range maps or GeoLifeCLEF presence–absence data, and compare it directly with SINR? These tasks also test generalization across data sources (from noisy presence-only observations to independent expert data), a complementary axis to CoordBench's spatial blocking. Conversely, do you expect the scale-adjustable representation to help species range mapping in poorly sampled regions? This seems like a natural point of contact between the socioeconomic and ecological mapping communities.
  4. Given that MINDSET is sampled only inside and around cities, how does MIND perform on rural and non-US targets specifically? Does the Equal Earth discontinuity at the antimeridian produce artifacts in the released grid (e.g., eastern Russia, Alaska, Pacific islands)? (see also "weaknesses" section above)
1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for reading the paper so closely. We answer Q1 under W1, Q2 under W4, and Q3 and the rural part of Q4 under W2.

Q4: … Does the Equal Earth discontinuity at the antimeridian produce artifacts …?

Yes, the encoder has a seam at ±180° because its Equal Earth input is discontinuous there, as is GeoCLIP's, which uses the same projection. In the released global embedding grid, we blend cells within 2.56° of the antimeridian with their mirror cells so the two sides meet (App. A). Downstream users working near ±180° can read the grid instead of running the encoder.

W1: Missing PCA baseline. …

We added PCA of all four teachers, including AEF (App. I, Table 12). With 64 components, PCA scores 0.289 R2R^2 at 10° and −0.253 at 40°, compared with 0.309 and 0.128 for MIND-64. When tuning the truncation using inner CV, PCA scores 0.268 vs. 0.321 for MIND at 10° and −0.200 vs. 0.134 at 40° (Tables 12 and 4). We also agree that the spectral-bias explanation is not tested, and Sec. 5.2 now presents it as a hypothesis.

W2: Generality vs. evaluation scope. …

Species (App. H, Table 11). MIND+CP scores above SINR on IUCN (0.734 vs. 0.661 mean AP at 10°) and eBird S&T (0.832 vs. 0.774 at 10°), and a MIND variant is best on every GeoLifeCLEF split. For your question about poorly sampled regions, the Chunked Penalty improves over MIND-Full only under regional holdout (IUCN: 0.937 for both under random folds, 0.734 vs. 0.720 at 10°). This is consistent with the penalty helping where nearby labels are absent.

Rural and non-US (App. J). When defining rural points as those more than 50 km from any city of at least 15k people, on the global WorldClim, SatCLIP, and air temperature datasets, MIND+CP is best or tied for best within both rural and urban groups, leads two cross-group comparisons outright, and ties for first in a third. It also has the highest R2R^2 at every block size on the 14 global or non-US regression datasets (0.583 vs. 0.502 for GeoCLIP at 10°).

W3: Incomplete comparison with the teachers. …

The general response above gives the direct AEF comparison (App. G). MIND+CP also exceeds the concatenation of the other three teachers, e.g. 0.508 vs. 0.496 R2R^2 at 2° (Table 12). Adding AEF to the concatenation has higher accuracy and R2R^2 under random folds and 2° blocks, however, MIND+CP has higher R2R^2 from 10° on (0.356 vs. 0.291 at 10°). The raw concatenation has no ordering by scale so we instead applied truncation and the Chunked Penalty to its PCA, with components ordered by variance (W1). With the Chunked Penalty, PCA scores −0.077 at 10° and −0.494 at 40°, compared with 0.356 and 0.184 for MIND+CP.

W4: The regional advantage comes from environmental targets …

We agree. Sec. 5.1 now notes that GeoCLIP has higher socioeconomic R2R^2 at 20° and 40°. The abstract no longer refers to generalizing across larger geographic gaps. Sec. 4 clarifies that the splits withhold downstream labels, while the encoders are pretrained on globally distributed data. For Q2, we think that because MIND has to reconstruct all four teachers it keeps a compromise among them and does not retain all of GeoCLIP's regional socioeconomic signal.

W5: Modest benefit of nesting, none for classification. …

We agree that the benefit of nesting is specific to regression and have made this clear in Sec. 5.3. Table 4 gives the comparison in the main text. When we pair results by fold assignment, MIND+CP scores above the non-nested model with standard ridge, its best setting, in all five fold assignments at both 10° and 40° (mean differences 0.018 and 0.057).

W6: Mean R^2 weights datasets equally …

Yes, that is how the weighting works, and we have made it explicit in App. F. Giving each source in Table 1 equal weight instead still leaves MIND+CP best at every block size (0.443 vs. 0.382 for GeoCLIP at 20°). Fig. 4 shows the results by category in the main text, and MIND+CP is ahead of GeoCLIP in 24/25 comparisons across splits and seeds and of Climplicit in 25/25 (Sec. 5.1). We also added macro-F1 and balanced accuracy for the land cover datasets (App. F, Table 7), and MIND+CP is highest or tied for highest in every split under both.

W7: Weak spatial baselines. …

Prior work rarely evaluates against spatial interpolation baselines, but we include them for transparency, even when they outperform our model on random folds. We added kriging (Table 3, App. L), which now has the highest random-fold R2R^2 and matches MIND+CP's random-fold accuracy (67.7% vs. 67.8%). From 10° on, it trails MIND+CP by 0.18–0.25 R2R^2 and 3.2–4.7 accuracy points. Sec. 2 now cites the area of applicability and design-based validation alongside distance-matched folds (NNDM).

W8: Novelty of the evaluation. …

We agree that SINR's Geo Feature task should be acknowledged and now cite it in Sec. 2. CoordBench adds global coverage, 78 targets, and block sizes up to 40°.

Reviewer_tdYAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of learning embedding for coordinates. Within this it introduces Matryoshka Implicit Neural Distillation (MIND), a technique which describes an embedding increasing with granularity in the higher dimensions. This embedding later on can be used for downstream tasks. They further introduce the dataset CoordBench, with the concept of regional holdouts at multiple spatial scales.

Critical strengths
  • A valuable comparison against a lot of different models.
  • Well presented and interesting topic. Could be a good starting point for future benchmarking in this field.
Critical weaknesses
  • I am wondering why you add the downstream prediction part in Figure 1 when not mention it in the text part.
  • All CoordBench datasets in the “Land Cover” category originate from a single source, the AEF evaluation suite (Table 1), and the accuracy results in Table 3 are identical to this category (Table 7). AEF is also one of MIND’s teachers, yet it is excluded both as a baseline (as a non-coordinate-only model) and from the teacher-concatenation control. It remains unclear how much of MIND’s advantage over the other INRs, particularly on Land Cover, is inherited from AEF rather than gained by MIND itself. I would like to see the benchmark results against AEF if possible, you of course can exclude it with an asterisk as it is not a full fair comparison in terms of data availability. Further, you have four teacher models, are all of them needed? Leave one teacher out or single teacher ablation is missing.
Questions for authors
  • I see you have used your MIND-Encoding and compared it against a variety of Domain Models (e.g. SINR ... SLED). Do the Domain Models exceed the performance of MIND on their specific task (you currently resolve it only in the supplementary according to the three groups (Environmental; Land Cover; Socioeconomic))? Is there a comparison where the Domain Model were only evaluated on the Test splits of their domain Task against MIND and MIND only trained on the splits without leakage?

  • Do you have a clear resolution where you show which Dataset and Measurements the Benchmark covers?

    I only found: "The benchmark suite covers socioeconomic targets such as income and housing (26 datasets), environmental quantities such as temperature and soil pH (10), and land-cover classes (16)"

    Give me an detailed table in the appendix. Cite the Dataset give me an Detailed list. Which target was involved in which category and so on.

  • Did you take into account the time splits ? As a lot of those data should be effected when the city increase e.g. the environment deteriorates. How do you cope with this dimension ?

  • Are those dataset used by any of the competitors, can you identify leakage ? Please make this as matrices, in the appendix as well. Further it may be important to mention when they are leakage free or the dataset is not clearly stated from the competitor. It would also be in general good to state which tasks the models are learned.

  • Which released checkpoint (variant, resolution, embedding dimension) was used for each baseline, nor whether any baseline and teacher? Could you please be more concrete if possible ? Please add an appendix table listing for every compared model the exact checkpoint, its source, its embedding dimension and state explicitly that all baselines are used with frozen released weights.

  • How the Metrics are calculated? Please detail it out latest in appendix? I want especially how the averages were calculated.

Especially the weaknesses and the question raised currently let me then to weak reject, but I would be glad to accept this paper if the rebuttal delivers the answers and increases the paper overall quality.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for the thorough review. We have added the requested comparisons and documentation to the revised PDF.

Q1: … Do the Domain Models exceed the performance of MIND on their specific task …?

On SatCLIP's own regression datasets (air temperature, elevation, population), MIND+CP scores 0.887 R2R^2 under random folds and 0.718 at 40°, compared with SatCLIP's 0.728 and 0.483. On WorldClim, the corresponding scores are 0.906 and 0.716 for Climplicit, and 0.950 and 0.829 for MIND+CP. MIND+CP also scores above SINR on species range maps (0.734 vs. 0.661 mean AP on IUCN at 10°; App. H). Domain models do score higher in some cases, for example, AEF on the 13 land-cover datasets from its own evaluation suite (75.4% vs. 62.2% accuracy under random folds; Table 10), and GeoCLIP on socioeconomic targets at 20° and 40° (0.203 vs. 0.165 and 0.038 vs. 0.003 R2R^2; Table 8).

Q2: … which Dataset and Measurements the Benchmark covers? …

We have added these details in App. M. Table 17 lists all 52 datasets and 78 targets by category (environmental, land cover, socioeconomic), with task type, sample count, coverage, spatial unit, and source. Table 18 lists the 27 PDFM targets, and Table 19 contains each source's citation, release, and license.

Q3: Did you take into account the time splits ? …

In this paper we don't study spatiotemporal INRs and all the holdouts we use are spatial. MIND and the large majority of the INRs we compare with take only coordinates as input. We agree that temporal change matters and have added temporal holdouts to the future work in Sec. 6.

Q4: … can you identify leakage ? Please make this as matrices, in the appendix as well. …

We've added this matrix to Table 20 (App. N) along with each model's pretraining data and objective/domain. It distinguishes supervision on the same variable from related domains or proxies, and marks datasets from a model's own evaluation release, such as SatCLIP's and AEF's tasks. We also clarify when we could not confidently identify the overlap. All encoders including MIND are frozen during evaluation so CoordBench labels are never used to train the encoder itself, only to fit the linear probes on each model's embeddings. The new TerraNova baseline addition is the only model that was explicitly pretrained on data products from CoordBench (two of the SoilGrids targets).

Q5: Which released checkpoint … was used for each baseline …?

We added Table 21 (App. N) which lists the released weights or variant, source, library/loader, and embedding dimension for every model, and explicitly states that all encoders use frozen released weights. The SINR, Climplicit, and GeoCLIP baselines use the same checkpoints that MIND distills on. All MIND rows use one released checkpoint.

Q6: How the Metrics are calculated? …

App. F now adds the equation and details for the CoordBench metric scoring. For each target, we compute R2R^2 on each held-out fold, average the five folds, and floor that average at −1. We then average targets within each benchmark dataset, weighting the datasets equally, and average over 5 fold-assignment seeds. Accuracy is averaged the same way but without flooring. As a sensitivity check, App. F also weights each regression source in Table 1 equally, and MIND+CP still has the highest R2R^2 at every block size (e.g. 0.443 vs. 0.382 for GeoCLIP at 20°).

W1: … why you add the downstream prediction part in Figure 1 when not mention it in the text part.

We have made the connection clearer in Sec. 3 which now points to Fig. 1b when describing how we fit the downstream predictor after discarding the reconstruction heads.

W2: All CoordBench datasets in the “Land Cover” category originate from a single source … Leave one teacher out or single teacher ablation is missing.

We agree that a direct AEF comparison is useful and have added it in App. G. Of the 16 land-cover datasets, 13 come from AEF's evaluation suite and the other three are SatCLIP's country, ecoregion, and biome tasks. AEF does better on its own 13 land-cover datasets while MIND+CP does better on the three SatCLIP datasets and on regression at every block size (0.356 vs. 0.287 R2R^2 at 10°). Table 12 compares MIND with each individual teacher and with their concatenations. MIND+CP leads the coordinate-only rows and from 10° on it has higher regression scores than all rows that require downloading AEF raster.

We also retrained MIND by leaving out one teacher at a time. Removing Climplicit, GeoCLIP, or SINR lowers random-fold R2R^2 by 0.029–0.038 compared to the 4-teacher model, while removing AEF changes it by 0.001. Across all block sizes and five seeds, removing AEF changes R2R^2 by at most 0.034 and land-cover accuracy by at most 0.6 points, so MIND's land-cover performance does not depend mainly on AEF. The Chunked Penalty also improves regional R2R^2 for every teacher set.

Reviewer_tdYAOpenReview ↗

Response to the Rebuttal

Dear authors, thank you for addressing my concerns. The presented rebuttal sounds feasible and I will gladly change my rating to weak accept.

Thank you again for the thoughtful review which helped improve the paper and for taking the time to engage with our rebuttal. Just a gentle flag that the rating in your official review doesn't appear to be updated yet.

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

Accurate modelling of environmental and socioeconomic factors globally remains a key challenge, especially in low-resource/low-data locations. Location encoders provide a promising approach to combat this issue. This paper proposes MIND - a multi-teacher distillation approach intended to work across coarse and fine spatial scales, along with CoordBench a comprehensive (if somewhat biased) dataset benchmark marking a key contribution to the evaluation and analysis of spatial encoders.

Critical strengths

Strengths Comprehensive dataset and empirical analysis, highlighting the importance of careful consideration of spatial splits. Strong results for the proposed method under the benchmark settings High impact research direction for disciplines beyond AI/ML

Critical weaknesses

Weaknesses The performance improvements seem somewhat reliant on Chunked Penalty. The method seems like an incremental improvement on existing work cited by the authors, so the novelty seems somewhat limited however, this is partially offset by the empirical work and the dataset contribution. Given the evaluation protocol, I have concerns regarding fair evaluation (CP tuning compared to other baselines) and leakage (Teachers vs test set overlap).

Questions for authors

Questions: How many times was MIND retrained? can you provide indications on the variance of the method across different runs. A broader discussion on the variance as a consequence of the holdout strategies and choice of teacher(s) would be good to see.

How are leakage concerns addressed? How much overlap do the teachers have with the test sets?

How much do the other baseline methods benefit from CP if their dimensions are first ordered?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Thank you for your time reviewing our paper and for recognizing the benchmark and empirical analysis. This review became available to us late in the rebuttal period, with only two days remaining, limiting our ability to complete additional experiments before the deadline. Below, we summarize the relevant evidence from the revised manuscript and experiments completed during the rebuttal. We answer W3 under Q2 and Q3.

W1: The performance improvements seem somewhat reliant on Chunked Penalty.

We agree that CP contributes to the regional gains. MIND also performs well without CP using standard ridge regression. MIND-Full has higher random-fold R2R^2 (0.642) and accuracy (67.8%) than the pretrained INR baselines. MIND-128, truncated and evaluated with standard ridge, scores 0.327 R2R^2 at 10° and 0.303 at 20°, above every baseline at full width (GeoCLIP is next with 0.303 and 0.287; Table 3). At 40°, GeoCLIP scores higher than MIND-128 (0.151 vs. 0.136).

W2: The method seems like an incremental improvement on existing work …

MIND combines nested multi-teacher distillation with an observed coarse-to-fine spatial ordering that allows scale to be adjusted after training. Sec. 2 discusses SLED, TerraNova, and many other INRs which don’t have this capability. Our controlled experiments also support the role of nesting in spatial ordering (Sec. 5.2, App. D) showing that the characteristic distance falls from 790 km for the first chunk to 200 km for the full embedding, but stays near 230 km at every truncation for the same architecture trained without nesting. CP uses this ordering through weighted ridge regression to provide even greater performance gains (Sec. 3.2).

We also release all code, checkpoints, CoordBench, and a global 1 km embedding grid, free to use with no egress fees.

Q1: How many times was MIND retrained? … variance … holdout strategies and choice of teacher(s) …

We used one released encoder for all main results. During evals we vary the folds across 5 seeds while keeping the encoder frozen. To check pretraining variability we reran pretraining with two more random seeds. At every regional block size, MIND+CP varies less across pretraining seeds than across fold assignments. Notably, we found that our released encoder is also the weakest of the three runs with the two new runs producing even stronger results for MIND.

MIND+CP's fold-assignment standard deviation increases from 0.006 R2R^2 at 10° to 0.076 at 40° (Table 9). We also test buffered exclusion from 10 to 100 km around fixed test points (App. E). For teacher choice, Table 12 compares individual teachers and their concatenations. We also retrained MIND leaving out each teacher. Removing Climplicit, GeoCLIP, or SINR lowers random-fold R2R^2 by 0.029–0.038, while removing AEF changes it by 0.001.

Q2: How are leakage concerns addressed? How much overlap do the teachers have with the test sets?

Table 20 (App. N) provides details on each teacher’s supervision, relevance to CoordBench target variables, and tasks from each model's evaluation release. For example, Climplicit reconstructs CHELSA climate variables, which are related to the WorldClim targets in CoordBench. Also, 13 of the 16 land-cover datasets come from AEF's evaluation suite. All encoders are frozen during evaluation, with probe fitting and tuning confined to outer training folds. These splits withhold downstream labels but do not establish that regions were unseen during pretraining (Sec. 4). Geographic INRs aim to fit a smooth function of the Earth from global data but are mostly evaluated with random folds, so CoordBench adds regional holdouts (Table 3). Removing AEF changes land-cover accuracy by at most 0.6 points, suggesting limited dependence on that teacher.

Q3: How much do the other baseline methods benefit from CP if their dimensions are first ordered?

We evaluated CP on variance-ordered PCA of the four teachers and two coordinate encodings with explicit coarse-to-fine ordering (App. I). For PCA, CP scores below CV truncation at every regional block size (−0.077 vs. 0.268 R2R^2 at 10°; Table 12). Ordering by variance does not guarantee spatial ordering. CP improves spherical harmonics ordered by degree (0.108 to 0.206 at 10°). For multiscale Fourier features ordered by length scale, CP does not improve R2R^2 at 10° or larger blocks (Table 13). MIND+CP scores 0.356 at 10°. GeoCLIP, non-nested MIND, and permuted MIND show no CP gain (Table 4), but GeoCLIP's dimensions were not reordered. We have not tested spatially ordering of every pretrained baseline.

For fairness, all representation probes use the same outer folds and inner validation splits (Sec. 5). CP tunes the ridge penalty and chunk growth, while truncation tunes the ridge penalty and dimension. Both use nested validation, and CP's candidates include standard ridge (App. A).

We are happy to clarify any of these points further.

Reviewer_sDSDOpenReview ↗

AEF as a teacher

Thank you for the reply, I am still considering several aspects of this work and the reply.

regarding your response - "while removing AEF changes it by 0.001".

Can you clarify why AEF is a teacher in the first place?

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Rebuttal Response by Authors

Thank you all for the time you put into reading the paper and suggesting improvements. The revised paper adds baseline comparisons with two satellite embedding products, two TerraNova encoders, two fixed coordinate encodings, and ordinary and indicator kriging. We also added evaluations on three species-distribution datasets, teacher ablations, and five benchmark documentation tables. The appendices are now in the main paper PDF. All CoordBench comparisons use the same folds, and we fit linear probes on frozen features for every embedding model.

  • AEF and TESSERA v1.1 (App. G, Table 10). Several reviewers asked how MIND compares with its teacher AEF on CoordBench. We evaluated both AEF and TESSERA v1.1, 10 m embedding products that require satellite image time series at prediction time, whereas MIND needs only coordinates. Both are more accurate than MIND+CP on the land-cover datasets from AEF's evaluation suite, which are labeled at the pixel level. At 10°, mean accuracy on these tasks is 72.1% for AEF and 69.9% for TESSERA, compared with 54.9% for MIND+CP. On the three tasks labeled by polygon (country, ecoregion, and biome), MIND+CP averages 80.0% accuracy at 10°, compared with 60.5% for AEF and 54.6% for TESSERA. MIND+CP also has higher mean regression R2R^2 than both products at every block size, including 0.356 vs. 0.287 for AEF at 10° and 0.184 vs. 0.047 at 40°.

  • Species distribution (App. H, Table 11). As reviewer 5pyS pointed out, CoordBench does not currently include species-distribution datasets, although one of MIND's teachers, SINR, specializes in this domain. We therefore added evaluations on the IUCN range maps, eBird Status & Trends, and GeoLifeCLEF datasets. MIND+CP scores above SINR in every split of all three datasets and so does MIND-256 at SINR's 256 dimensions.

  • TerraNova (Table 3, Sec. 2). We thank reviewer KSK9 for suggesting this concurrent work. We were unaware of it at submission but are happy to include the comparison. We added both TerraNova encoders (spatial and spatiotemporal) as baselines and cite the paper. TerraNova is a strong model among other INRs. MIND+CP scores above both of its encoders and so does MIND-256 at TerraNova's 256 dimensions.

  • Teachers (App. I, Table 12). MIND+CP exceeds each coordinate-only teacher and their concatenation on both metrics in every split. Adding AEF to the concatenation gives higher accuracy in every split and higher R2R^2 under random folds and 2° blocks, but from 10° on MIND+CP has higher R2R^2 than all four teachers combined, including after PCA with tuned truncation or the Chunked Penalty. At 10°, R2R^2 is 0.356 for MIND+CP, 0.291 for the concatenation with AEF, and 0.259 without it. We also retrained MIND leaving out each teacher in turn (reported in the individual responses). Removing AEF changes land-cover accuracy by at most 0.6 percentage points across all block sizes.

  • Frequency-ordered encodings (App. I, Table 13). We tested spherical harmonics up to degree 40 and multiscale random Fourier features. We evaluated both with the same probes as MIND (standard ridge, truncation with a tuned frequency cutoff, and the Chunked Penalty) to ensure a fair comparison. Results show that MIND+CP scores above both encodings in all cases.

  • Additional spatial baselines (Table 3, App. L). We added ordinary kriging for regression and indicator kriging for classification alongside the existing inverse distance weighting (IDW) baseline. Under random folds, ordinary kriging and IDW have higher regression R2R^2 than MIND+CP (0.685 and 0.675 vs. 0.643). MIND+CP leads both methods at every regional block size in Table 3. At 10°, it reaches 0.356 R2R^2, compared with 0.175 for ordinary kriging and 0.187 for IDW. On classification at the same block size, MIND+CP reaches 59.9% accuracy, compared with 56.7% for indicator kriging and 55.7% for IDW. Reviewer 5pyS also asked for macro-F1 and balanced accuracy on land cover (App. F, Table 7). Under both, MIND+CP scores above IDW and indicator kriging in every split, including random folds (46.2 vs. 45.9 macro-F1 and 45.3 vs. 45.1 balanced accuracy for indicator kriging).

  • Benchmark documentation (App. M–N, Tables 17–21). The appendices now list each dataset by category, with provenance and licenses, and provide a matrix of pretraining overlap and the exact released checkpoint we used for every model in the paper. The code, benchmark, and checkpoints are in the anonymized supplementary material (see the Reproducibility Statement)

  • Other additions. We report urban versus rural and non-US results (App. J), intrinsic dimension (App. K), class-balanced land-cover metrics, and a check with aggregation by source (App. F).

Area_Chair_LfXqOpenReview ↗

Author Reviewer Discussion

Dear reviewers,

Thanks again for agreeing to review for ICLR.

As we are approaching the end of the discussion period, I would appreciate your acknowledgment of the authors' rebuttal and any feedback you would like to share with the authors.

This will help to improve the submissions, streamline decision making, and keep ICLR's original idea of having such a discussion period valuable to all participants in the review process.

Best, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.