{"id":"b33e94d2-6cd4-49c8-817f-95a552284f62","arxiv_id":"2502.06831","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Spherical wavelet encodings reduce the performance gap on small and coastal landmasses that spherical harmonic and other location encodings exhibit in implicit neural representations of Earth data.","lead":"This paper introduces FAIR-Earth, a benchmark dataset with geographic and demographic metadata for measuring how fairly implicit neural networks model Earth signals, and shows that islands and coastlines are consistently predicted worse than large landmasses. It then proposes a spherical wavelet location encoding that improves these subgroups' performance while matching global accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'sharp contrast' in local-global correlation rests on four resolutions with post hoc exclusion of SPHERE C+; SW correlations are non-monotonic (0.51, 0.42, 0.04, 0.31), so the positive-trend claim is not yet established.","rationale":"The reader's weakest_assumption focuses on polar degradation of the spherical wavelet encoding, which is acknowledged and precisely located in Fig. 23 and Section 4.3's tradeoff paragraph. While this is a real limitation, it does not directly undercut the specific claim about islands and coastlines; it is a boundary condition on the method's name rather than the core empirical comparison. The more load-bearing issue is the fragility of the correlation evidence that supports the central claim of a 'sharp contrast' between SW and baselines. Table 2 supplies only four resolutions, no uncertainty quantification, and a post-hoc exclusion of SPHERE C+. Given that SW's own correlations fluctuate from 0.51 to 0.04 across those four points, the claim of a consistent positive trend is not statistically anchored. This is not an accusation of cherry-picking; it is a statement that the paper's headline fairness result is under-determined by the presented numbers. A straightforward extension of the same experiment (more resolutions, all baselines, bootstrap/permutation tests) would settle it. The dataset contribution and the qualitative figures still support a conditional acceptance, so I would not move the verdict to REJECT; I agree with CONDITIONAL. My read partially overlaps with the reader's rationale, which also flagged the SPHERE C+ exclusion, but I do not think the polar issue is the single most load-bearing concern.","tokens_in":17881,"tokens_out":3665,"duration_ms":28914,"concrete_test":"Re-run the land-island correlation experiment on at least 8 training resolutions (e.g., 2500, 5000, 7500, 10000, 12500, 15000, 17500, 20000) with all baselines including SPHERE C+, and report each Pearson r with a bootstrap 95% CI or permutation p-value, plus a leave-one-resolution-out sensitivity table. If SW correlations are not consistently positive, or are not significantly higher than SH and THEORY after including SPHERE C+, the 'sharp contrast' claim in Section 4.3 should be withdrawn or substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 and Table 2 base the central fairness claim on Pearson correlations between land and island loss across training resolutions. There are only four resolutions (5000, 10000, 15000, 20000), so each r is estimated with n=4, and no confidence interval, standard error, or significance test is reported. The SW correlations are 0.51, 0.42, 0.04, 0.31; this sequence is not consistently positive and includes a near-zero value. SPHERE C+ is excluded from Table 2 'due to abnormally high bias', a post-hoc exclusion that weakens the baseline comparison. With only four points, a single resolution (15000, where SW r=0.04) can flip the qualitative conclusion, and leave-one-out sensitivity is unknown. The claimed 'positive trend between global (land) and localized (island) performance' is therefore not robustly supported by the presented evidence. Because this correlation result is the main empirical support for SW's fairness advantage over baselines, the central claim is under-determined as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces FAIR-Earth, an open-source benchmark package for measuring subgroup-level performance of implicit neural representations (INRs) on Earth data, with metadata stratifications for islands, coastlines, population density, and countries. Using this benchmark, the authors evaluate spherical harmonic, theory, grid-and-sphere, and their proposed Spherical Wavelet (SW) encodings, and report that existing encodings show strong biases against islands and coastlines. They propose SW encodings built from inverse stereographic projections of Morlet wavelets over a Fibonacci lattice, and argue that SW produces positive correlations between land and island performance across training resolutions, reduces island/coastline losses on the land-sea task, and remains competitive with spherical harmonics on global metrics. The paper also includes external validation on checkerboard and Rußwurm land-sea tasks and reports computational comparisons.","tokens_in":18112,"tokens_out":9568,"duration_ms":89069,"significance":"The FAIR-Earth dataset and accompanying open-source package are a useful contribution: they provide a concrete way to stratify Earth INR evaluation by geographic and demographic subgroups, and the empirical documentation of island/coastline bias in spherical harmonic and related encodings is valuable and credible. The SW encoding is a reasonable proposal, and the paper includes external tasks, hyperparameter sweeps, and error analyses that go beyond a single benchmark. However, the central claim that SW 'corrects localized biases' is currently supported mainly by the land-sea classification task; the correlation evidence in Table 2 is under-specified, and the temperature subgroup results in Table 15 show SW worse than SH on every subgroup. The theoretical claim of an orthogonal wavelet basis is also not established. With these points clarified, the paper could be a solid contribution.","major_comments":[{"comment":"The claim that SW 'induces a positive trend between global (land) and localized (island) performance' rests on Pearson correlations whose sample size is not reported. If each correlation is computed across hyperparameter configurations at a fixed training resolution, the number of pairs needs to be stated and confidence intervals or significance tests provided; if it is computed across the four resolutions, n=4 and the sequence 0.51, 0.42, 0.04, 0.31 is not consistently positive and the conclusion would be fragile to a single resolution. The post hoc omission of SPHERE C+ from Table 2 'due to abnormally high bias' should be replaced by a prespecified exclusion criterion or by an appendix table including all baselines.","section":"Section 4.3, Table 2"},{"comment":"The text states that the rotated and dilated Morlet wavelets form 'a set of orthogonal basis functions.' Inverse stereographic projection of an admissible Euclidean wavelet yields an admissible spherical wavelet (zero-mean), but admissibility does not imply orthogonality, and Morlet wavelets are typically frames rather than orthogonal bases; discretization over a Fibonacci lattice with dyadic dilations does not by itself restore orthogonality. Please either prove the claim, replace it with the correct frame terminology, or provide a numerical check (e.g., Gram matrix condition) if orthogonality is used in the reconstruction formula.","section":"Section 4.2, after Eq. (2)"},{"comment":"The paper acknowledges that the inverse stereographic construction is ill-defined at the poles and that SW performance degrades at high latitudes. Because the title and framing promise that 'no location is left behind,' the paper should quantify the affected area or population and state explicitly that the fairness improvement applies away from polar regions. As written, the polar degradation is a known limitation but its scope is not assessed.","section":"Section 4.3 and Fig. 23"},{"comment":"On the surface temperature regression task, SW has higher subgroup loss than SH for every subgroup (land 0.087 vs. 0.076, sea 0.043 vs. 0.028, island 0.049 vs. 0.041, coast 0.101 vs. 0.083). Thus the claimed fairness improvement over baselines is demonstrated only on the land-sea binary classification task, not across the modalities promised in the abstract. The paper should either restrict the fairness claim to the land-sea task or provide subgroup-level evidence showing improvement on other FAIR-Earth modalities.","section":"Section 4.3 and Table 15"}],"minor_comments":[{"comment":"The sea loss at resolution 30000 is reported as 0.80 ± 0.03, while all other sea entries are 0.16 or below; this appears to be a typo for 0.08 and should be corrected.","section":"Table 1"},{"comment":"The phrase 'data specifications available in Appendix A.1 and Appendix A.1' should reference the correct appendix sections, as the duplicate reference is a typo.","section":"Section 3.2"},{"comment":"The sentence 'particularly struggle with to demarcate the Spain's fine Mediterranean coastline' contains a grammatical error and should be rephrased.","section":"Section 4.1"},{"comment":"The description of the 0.1° grid as 'state-of-the-art resolution' is an overstatement for some of the constituent datasets and should be qualified.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The work is likely of interest to the geo-AI and INR communities, and the FAIR-Earth package is a real contribution. However, the central fairness claim currently rests on a single task and an under-specified correlation analysis. I would encourage a revision that tightens the claims, reports uncertainty for the correlation results, and addresses the orthogonality and polar-degradation issues rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful paper, more for the FAIR-Earth benchmark than for the spherical wavelet encoding as such. The dataset fills a real gap: a standardized set of Earth signals with subgroup metadata for testing location-based bias in INRs. The wavelet encoding is a reasonable new combination, and the experiments on land-sea and temperature tasks do show competitive or better subgroup losses relative to spherical harmonics. I would use the dataset; I am less sure about the strength of the central fairness claim.\n\nThe main empirical support for 'no location left behind' is the correlation analysis in Table 2. The stress-test's claim that each correlation uses only four points is not correct; these are correlations across hyperparameter configurations within each resolution, not across the four rows. But the analysis is still under-reported. No sample size, no confidence interval, no significance test for any r. And the SW values are 0.51, 0.42, 0.04, 0.31—hardly a smooth 'positive trend' in any direction. The post hoc omission of SPHERE C+ from Table 2, justified only by 'abnormally high bias', would be more acceptable with a clear criterion stated in advance.\n\nOther soft spots: the phrase 'orthogonal basis functions' in Section 4.2 is not backed by the construction; these are frames, and the paper doesn't need orthogonality for anything downstream, so it's sloppy but harmless. Table 1 has a likely typo (SEA loss 0.80 at 30000 resolution vs 0.08 at 25000), which makes me want to inspect the released data. The polar degradation is acknowledged and visibly large in Fig 23, yet the title and abstract say 'no location left behind'; that is an overclaim.\n\nOn the positive side, the paper is honest about the tradeoffs: SW is worse on the checkerboard task, and the fine-tuning results show it is roughly on par, not uniformly better. The external validation on Rußwurm's land-sea task is a good check. The code and data are promised open source, which is exactly what the community needs for a fairness benchmark.\n\nWho is this for? Anyone working with location encodings for Earth data, and anyone who wants a standard fairness evaluation for geospatial INRs. I would send it to a serious referee. My recommendation would be a conditional accept, with requests for confidence intervals on the correlations, a pre-registered or principled exclusion rule for baselines, and a more careful title.","headline":"A useful fairness benchmark for Earth INRs, with a plausible wavelet mitigation that is oversold by the title and a correlation analysis that needs more statistical care.","tokens_in":18630,"tokens_out":4965,"would_cite":true,"duration_ms":44504,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Spherical wavelets end Earth models' island blind spot.","keywords":["implicit neural representations","Earth data fairness","spherical wavelets","spherical harmonic encodings","location encoding","subgroup performance disparity","multi-scale representation","benchmark dataset"],"falsifier":"On a FAIR-Earth task, restrict evaluation to latitudes above $80^\\circ$ and compare spherical-wavelet and spherical-harmonic error; if the wavelet encoding's polar error exceeds the harmonic encoding's by more than the island error dropped, the claim that no location is left behind is not supported for polar regions.","tokens_in":17644,"feed_emoji":"🌍","tokens_out":9004,"duration_ms":77223,"temperature":0.7,"pith_summary":"Earth-data implicit neural representations (INRs) are usually judged by global average error, which hides where they fail. This paper claims that those failures are systematic: islands, coastlines, and other high-frequency, localized regions are consistently modeled worse than large landmasses, and the global spherical-harmonic bases used by state-of-the-art encodings force a tradeoff between global and local accuracy. To make the problem measurable, it introduces FAIR-Earth, a benchmark dataset with land-sea, temperature, precipitation, CO2, and population signals, plus metadata for stratifying by landmass size, coast distance, population density, and country. To fix the bias, it proposes spherical wavelet encodings, which use multi-scale Morlet wavelets on the sphere and show positive land-island correlation with competitive global performance. The paper's caveat is that the wavelet construction is ill-defined at the poles, where its error grows.","feed_headline":"Spherical wavelets end Earth models' island blind spot","feed_subtitle":"A new benchmark shows islands and coastlines are mis-modeled; wavelet encodings reconcile local and global accuracy.","key_machinery":"The load-bearing object is a spherical Morlet mother wavelet, $\\psi_M(\\theta,\\phi) = [\\Pi^{-1}\\psi_{\\mathbb{R}^2}](\\theta,\\phi) = e^{i\\tan(\\theta/2)\\cos\\phi}\\,e^{-\\frac{1}{2}\\tan^2(\\theta/2)}/(1+\\cos\\theta)$, obtained by inverse stereographic projection of the Euclidean Morlet wavelet. From it, the encoding builds an orthogonal family by rotations $R(\\rho)$ over a Fibonacci lattice $F_N$ and dyadic dilations $D(a)$ with $a \\in \\{2^{i/Q}\\}$, giving $O(NM)$-size embeddings that localize signals in both dilation and rotation. This multi-scale localization is what lets the encoding represent islands and coastlines at coarse resolutions without aliasing the global land-sea structure. The same construction is what breaks near the poles: the inverse stereographic projection is ill-defined there, and the paper's latitudinal analysis shows error rising from equator to poles.","core_discovery":"The paper's central claim is that optimizing Earth INRs for global loss is not fairness-neutral: on FAIR-Earth, spherical-harmonic, theory, and grid/sphere encodings all exhibit much higher loss on islands and coastlines than on land or sea, and when trained at finer resolution the harmonic encodings trade off land accuracy against island accuracy, producing a negative correlation. The authors attribute this to the global support of Fourier bases, which alias around sharp localized discontinuities. Their proposed spherical wavelet encoding replaces global harmonics with a multi-scale basis built from the inverse stereographic projection of the Morlet wavelet, discretized over a Fibonacci lattice with dyadic dilations; this encoding maintains competitive or better average loss while producing a positive correlation between land and island performance, meaning local and global accuracy improve together. The authors also report that spherical wavelets are computationally cheaper at large encoding sizes than spherical harmonics, though they degrade at high latitudes and underperform harmonics on a coarse single-scale checkerboard task.","pith_inferences":["If the multi-scale property transfers, combining a global harmonic branch with a localized wavelet branch could remove the polar degradation while keeping the island gain; the paper does not test such a hybrid.","Because FAIR-Earth's grid is uniform in latitude and longitude, polar cells cover much smaller areas than equatorial cells, so the fairness metrics may under-weight polar surface area and soften the polar caveat.","The same wavelet-encoding idea could be evaluated on other INR domains with multi-scale structure, such as 3D scene or medical signal reconstruction, to see whether the positive land-island correlation generalizes beyond Earth data.","The reported negative land-island correlation for spherical harmonics could be used as a cheap diagnostic in other geospatial benchmarks: compute per-subgroup loss on land versus small landmasses before deploying an Earth INR."],"forward_implications":["Spherical wavelet encodings turn the land-island performance correlation from negative to positive across training resolutions, so optimizing for global accuracy no longer sacrifices island accuracy.","On FAIR-Earth's land-sea, temperature, precipitation, and CO2 tasks, spherical wavelet encodings match or beat spherical harmonic encodings in average loss while narrowing island and coastline gaps.","Spherical wavelet encodings scale better in generation time at larger sizes than spherical harmonic encodings, which suffer factorial-related numerical instability at high Legendre degree.","The method does not dominate everywhere: spherical harmonics remain better on the coarse single-scale checkerboard task, and spherical wavelets degrade toward the poles.","FAIR-Earth enables country-level and population-stratified fairness audits that were not possible with previously available geospatial datasets."],"supporting_citations":[{"why":"Supplies the spherical-harmonic location encoding and SirenNet pipeline that the paper uses as its main baseline and improves upon.","marker":"Rußwurm et al. (2024)"},{"why":"Introduces the inverse stereographic projection construction of spherical wavelets that the spherical wavelet encoding is built on.","marker":"Demanet & Vandergheynst (2003)"},{"why":"Proves that inverse stereographic projection of an admissible Euclidean wavelet yields an admissible spherical wavelet.","marker":"Sanz et al. (2006)"},{"why":"Provides the rotation-dilation localization formulation and fast directional continuous spherical wavelet transform used in the encoding.","marker":"McEwen et al. (2007)"},{"why":"Supplies the Fibonacci lattice point distribution used to discretize rotations on the sphere.","marker":"Gonzalez (2009)"},{"why":"Supplies the dyadic dilation discretization and the wavelet/Fourier aliasing analysis motivating the multi-scale approach.","marker":"Mallat (1999)"},{"why":"Provides the IMERG land-sea data that forms the FAIR-Earth land-sea modality.","marker":"Huffman et al. (2014)"},{"why":"Provides the CHELSA temperature and precipitation data used for FAIR-Earth's environmental modalities.","marker":"Karger et al. (2017)"},{"why":"Supplies the GRID AND SPHERE baseline encodings compared in the fairness evaluation.","marker":"Mai et al. (2023)"},{"why":"Provides the sinusoidal representation network architecture used as the INR backbone with all encodings.","marker":"Sitzmann et al. (2020)"}],"fun_headline_variants":["Wavelets wipe out Earth models' island blind spot","Islands and coasts exposed as Earth AI's fairness gap","Spherical wavelets give islands a fair shot in Earth models","New Earth benchmark shames models on islands; wavelets fix it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The encoding is built from a projection that is not defined at the poles and that visibly degrades near them; the paper assumes those polar regions are a small enough share of Earth-data tasks that the fairness gain on islands and coastlines is not offset by a newly introduced high-latitude bias.","fun_headline_variants_meta":{"raw":{"variants":["Wavelets wipe out Earth models' island blind spot","Islands and coasts exposed as Earth AI's fairness gap","Spherical wavelets give islands a fair shot in Earth models","New Earth benchmark shames models on islands; wavelets fix it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000771,"raw_usage":{"total_tokens":3416,"prompt_tokens":951,"completion_tokens":2465,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2396}},"tokens_in":567,"tokens_out":2465,"duration_ms":16903,"temperature":1.0,"reasoning_tokens":2396,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:00:01.594485+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a FAIR-Earth task, restrict evaluation to latitudes above $80^\\circ$ and compare spherical-wavelet and spherical-harmonic error; if the wavelet encoding's polar error exceeds the harmonic encoding's by more than the island error dropped, the claim that no location is left behind is not supported for polar regions.","supporting_citations":[{"cited_title":"Gabor wavelets on the sphere","cited_arxiv_id":null,"evidence_quote":"Introduces the inverse stereographic projection construction of spherical wavelets that the spherical wavelet encoding is built on."},{"cited_title":"A wavelet tour of signal processing","cited_arxiv_id":null,"evidence_quote":"Supplies the dyadic dilation discretization and the wavelet/Fourier aliasing analysis motivating the multi-scale approach."}],"review_version":1}