Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Fusing spherical-harmonic location features with cleaned satellite images cuts childhood-poverty prediction error by nearly a fifth.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 10:07 UTC pith:HVUVLSK4

load-bearing objection Solid engineering upgrade to KidSat with real MAE gains and clean ablations; the 18.83% headline over-packages three interventions, but SH still helps and the work is worth refereeing. the 3 major comments →

arxiv 2607.08281 v1 pith:HVUVLSK4 submitted 2026-07-09 cs.CV stat.AP

Enhancing the KidSat Model: Integrating Geographical Encoding and Data Quality Assessment for Childhood Poverty Prediction

Hou Hin Ip , Ka Nam Lam , Joshua Man Yu Ng , Makkunda Sharma , Seth Flaxman , Codie Gerlach-Wood , H Juliette T Unwin This is my paper
classification cs.CV stat.AP
keywords satellite imagerychildhood povertyspherical harmonicsDINOv2image quality assessmentgeographic encodingDHS surveysAfrica
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Satellite-based poverty maps often fail because survey labels are sparse and noisy, many Landsat scenes are clouded or corrupted, and pure vision models ignore geography. This paper upgrades an existing childhood-poverty pipeline by shrinking a sparse one-hot fine-tuning matrix, systematically discarding the worst images, and concatenating DINOv2 visual embeddings with fixed spherical-harmonic encodings of latitude and longitude. The combined changes lower mean absolute error on the cluster-level severe-deprivation proportion from 0.2167 to 0.1759 (an 18.83 percent relative drop) across 16 African countries; the same recipe scaled to 33 countries reaches 0.1658. Spherical harmonics alone consistently help; a higher-capacity neural location encoder does not, while gradient-boosted trees best exploit the fused features. The result is a practical, fully public-data recipe for more accurate, continent-scale childhood-poverty maps.

Core claim

Refining the fine-tuning target, filtering degraded Landsat scenes, and fusing DINOv2 embeddings with spherical-harmonic location features reduces MAE on cluster-level severe-deprivation proportions from 0.2167 to 0.1759 (18.83 percent relative improvement) on the original 16-country set, and yields 0.1658 when the best configuration is extended to 33 African countries.

What carries the argument

Spherical-harmonic (SH) geographic encodings: fixed multi-scale features of latitude and longitude (degree L=15) that are concatenated with DINOv2 visual embeddings and fed to a regression head, most effectively a gradient-boosted tree.

Load-bearing premise

The fixed 30 percent pixel-degradation threshold (and the percentile cloud cut-offs) is assumed to remove only uninformative noise rather than systematically discard usable but cloudy scenes in humid or coastal regions.

What would settle it

Re-run the identical pipeline with the degradation threshold set to 10 percent or 50 percent (or with no filtering) and check whether the reported MAE gains disappear or reverse on the same held-out clusters.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper enhances the KidSat pipeline for predicting cluster-level childhood severe-deprivation proportions from Landsat imagery and DHS surveys. Three interventions are introduced: (1) re-aggregation of sparse one-hot DHS variables (plus two new predictors) that reduces the fine-tuning matrix from 103 to ~51 dimensions; (2) a two-stage Landsat quality filter (scan-line/null detection + FMASK-inspired cloud indices with 30 % degradation threshold) that replaces ~15 % of scenes; and (3) fusion of frozen DINOv2 embeddings with parameter-free Spherical Harmonics (L=15) location features, optionally passed through a pre-trained SIREN. Five-fold CV on matched partitions shows the image-only MAE falling from 0.2167 to 0.1980 after preprocessing/quality filtering, then to 0.1759 with SH + LightGBM (18.83 % relative reduction). The same configuration yields MAE 0.1658 when scaled from 16 to 33 African countries. SH consistently improves over the image-only backbone; SH+SIREN underperforms; tree heads outperform ridge and MLP.

Significance. If the reported gains hold under fuller attribution, the work supplies a practical, fully public-data recipe that measurably improves a published satellite-poverty benchmark and demonstrates geographic scalability across 33 African countries. The clean ablation of SH versus SH+SIREN, the public-code commitment, and the use of parameter-free spherical harmonics are genuine strengths that other remote-sensing socioeconomic models can adopt. The contribution is incremental engineering rather than a new theoretical principle, yet the magnitude of the MAE reduction and the multi-country extension make it useful for the applied community.

major comments (3)
  1. Table 1 and §4.1 present the headline 0.2167 → 0.1759 (18.83 %) reduction as the joint effect of the three interventions, yet the design is only partially factorial. The quality filter permanently replaces ~15 % of scenes before any encoder ablation, and the best result further switches from the original ridge head to LightGBM. Consequently it is impossible to isolate how much of the 0.0408 absolute drop is attributable to genuine geographic signal versus cleaner supervision, selection of less-cloudy scenes, or the change of head. A minimal 2×2 (or 2×2×2) factorial—image-only vs SH, original vs filtered imagery, ridge vs LightGBM—on the same folds is required before the strongest claim can be accepted at face value.
  2. §2.2 fixes the 30 % degradation threshold and the 85th/90th-percentile cloud cut-offs without sensitivity analysis. Because the filter is applied once and permanently, any systematic removal of informative but cloudy scenes (especially in humid or coastal regions) would inflate the reported MAE gains through selection rather than signal recovery. At minimum the authors should re-run the best configuration under a small grid of thresholds (e.g., 20–40 %) and report whether the ranking of SH versus image-only remains stable.
  3. §2.3 and Eq. (1)–(2): the SH+SIREN encoder is pre-trained on the identical poverty vector that supervises DINOv2 fine-tuning. The discussion (§5.2) correctly notes possible representation overlap, yet no alternative objective (auxiliary demographics, reconstruction, contrastive alignment) is tested. Without that control the claim that “higher-capacity coordinate MLPs can underperform without carefully designed objectives” remains an untested hypothesis rather than a demonstrated result.
minor comments (5)
  1. Abstract and §2.1 state the one-hot reduction as “103 to 51” while the body text says “103 to 48”; Appendix B should be made consistent with the abstract figure.
  2. Figure 3 caption and §4.2 claim “17 additional countries” while the abstract and Table C list an expansion from 16 to 33 (i.e., +17). Clarify the arithmetic.
  3. Appendix D lists SIREN ω0=30 and early-stopping on LR < 1e-7; these hyper-parameters should be mentioned briefly in the main text or a footnote so that the under-performance of SH+SIREN can be interpreted without consulting the appendix.
  4. The original KidSat MAE of 0.2167 is reproduced under the authors’ own re-implementation; a short note confirming that the same random seeds and fold partitions recover the published number would strengthen the baseline claim.
  5. Typographical inconsistencies appear in the title (“THEKIDSATMODEL”) and several section headings; a final copy-edit pass is needed.

Circularity Check

0 steps flagged

No significant circularity; purely empirical MAE gains on held-out folds with parameter-free SH features.

full rationale

The paper is an empirical computer-vision / ML study that reports measured reductions in cluster-level MAE under 5-fold cross-validation (train on 80 % of clusters, evaluate exclusively on the held-out 20 %). The Spherical-Harmonics encoder is a fixed, parameter-free mathematical transform of latitude/longitude; the DINOv2 backbone and regression heads are trained only inside each fold; image-quality filtering and DHS re-aggregation are deterministic preprocessing steps applied before any training. No equation, fitted constant, or self-citation is used to force the claimed MAE numbers (0.2167 o 0.1759, etc.). Self-citation of the original KidSat work merely supplies the baseline pipeline that is re-implemented and improved; the improvements themselves are independently measured. Consequently the derivation chain contains no self-definitional, fitted-as-prediction, or load-bearing self-citation circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central MAE claims rest on a handful of hand-chosen thresholds and architectural constants plus standard remote-sensing and survey assumptions; no new physical entities are postulated.

free parameters (4)
  • SH degree L = 15
    Spherical harmonics are truncated at L=15, producing a fixed-dimensional location vector; the degree is stated without cross-validation or sensitivity analysis.
  • image degradation threshold = 30%
    Scenes with >30% degraded pixels are replaced; the cut-off is chosen by the authors and directly controls the 15% replacement rate.
  • cloud-detection percentiles = 85th / 90th
    Adaptive thresholds set at the 85th percentile (Landsat 7) and 90th percentile (Landsat 8) for brightness/NDVI/whiteness indices.
  • SIREN architecture and training = 4×256, ω0=30
    Four hidden layers of width 256, ω0=30, 200 epochs, learning rate 1e-4; these control the learned geo-embedding that under-performs plain SH.
axioms (4)
  • standard math Real spherical harmonics Y_ℓ^m up to degree L form a complete multi-scale basis on the sphere and can be concatenated with visual embeddings without further learning.
    Invoked in Section 2.3 and Equation (1); standard representation theory.
  • domain assumption DHS cluster-level severe-deprivation proportions constructed from the six UNICEF dimensions constitute a reliable continuous supervision target on [0,1].
    Inherited from the original KidSat pipeline and used throughout fine-tuning and evaluation.
  • domain assumption Physics-based spectral indices (brightness, NDVI, blue-to-red, whiteness) plus morphological filtering adequately detect cloud and scan-line corruption for Landsat 7/8.
    Section 2.2 cites FMASK; the two-stage procedure is treated as ground truth for image replacement.
  • ad hoc to paper Informed manual re-aggregation of sparse DHS categorical codes (guided by codebooks and prior literature) preserves predictive information while reducing sparsity.
    Appendix B details the merges; no automated feature-selection baseline is provided to justify the exact groupings.
invented entities (1)
  • two-stage Landsat quality-screening procedure no independent evidence
    purpose: Systematically replace the first-available image with the lowest-degradation image per cluster.
    Combines scan-line/null-pixel detection with FMASK-inspired cloud masks; the exact 30% rule and percentile cut-offs are paper-specific.

pith-pipeline@v1.1.0-grok45 · 18870 in / 2947 out tokens · 49054 ms · 2026-07-10T10:07:22.339916+00:00 · methodology

0 comments
read the original abstract

Accurate poverty mapping using satellite imagery is often hindered by (i) noisy and sparse survey-derived supervision, (ii) image quality issues such as cloud cover and image corruption, and (iii) lack of explicit spatial structure in image-only models. Building on the KidSat framework, we develop an enhanced pipeline that improves predictive accuracy via refined data preprocessing, systematic image quality assessment, and mathematically defined geographic encoding. First, we refine the fine-tuning target matrix by resolving high-cardinality sparsity and reducing one-hot dimensionality from 103 to 51 via DHS re-aggregation. Second, we introduce a simple two-stage quality-screening procedure to filter heavily clouded or corrupted observations. Third, we fuse DINOv2 visual embeddings with Spherical Harmonics (SH) location features. Across extensive experiments, these changes reduce MAE from 0.2167 to 0.1759, corresponding to an 18.83% relative reduction on the cluster-level severe-deprivation proportion scale. When extended from 16 to 33 African countries, the best-performing configuration achieves an overall MAE of 0.1658. We find that SH features consistently improve performance over the image-only backbone, whereas higher-capacity coordinate Multi Layer Perception augmentation (SH+SIREN) can underperform without carefully designed objectives. Finally, gradient-boosted tree heads (XGBoost/LightGBM) most effectively exploit nonlinear interactions in the fused visual-geographic representation. These findings provide a scalable and principled recipe for improving satellite-based socioeconomic predictions using only publicly accessible data.

Figures

Figures reproduced from arXiv: 2607.08281 by Codie Gerlach-Wood, H Juliette T Unwin, Hou Hin Ip, Joshua Man Yu Ng, Ka Nam Lam, Makkunda Sharma, Seth Flaxman.

Figure 1
Figure 1. Figure 1: Three multi-band (RGB) images from Luanda, Angola (early 2015) captured at 16–18 day intervals. The left [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Extended KidSat Model Architecture. Left: DINOv2 is fine-tuned against the poverty vector. Right: the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualisation demonstrating comparison between Baseline method (Left) and visualisation with modified [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages · 8 internal anchors

  1. [1]

    Measuring coverage in mnch: Tracking progress in health for women and children using dhs and mics household surveys.PLOS Medicine, 10(5):e1001391, 2013

    Ayse Hancioglu and Fred Arnold. Measuring coverage in mnch: Tracking progress in health for women and children using dhs and mics household surveys.PLOS Medicine, 10(5):e1001391, 2013

    work page 2013

  2. [2]

    Gray, Peter Azzopardi, Elissa Kennedy, Elise Willersdorf, and Mick Creati

    Natalie J. Gray, Peter Azzopardi, Elissa Kennedy, Elise Willersdorf, and Mick Creati. Improving adolescent reproductive health in asia and the pacific: Do we have the data? a review of dhs and mics surveys in nine countries.Asia-Pacific Journal of Public Health, 25(2):134–144, 2013

    work page 2013

  3. [3]

    The dhs program: Demographic and health surveys

    The DHS Program. The dhs program: Demographic and health surveys. https://dhsprogram.com/, 2025. Accessed: 2025-11-11

    work page 2025

  4. [4]

    The multiple indicator clus- ter surveys (mics) programme of unicef [webinar]

    The National Centre for Social Research (NatCen). The multiple indicator clus- ter surveys (mics) programme of unicef [webinar]. https://natcen.ac.uk/events/ multiple-indicator-cluster-surveys-mics-programme-unicef , May 2024. Online webinar, 1 May 2024, 17:00–18:00. 8

    work page 2024

  5. [5]

    The end of the dhs programme: A major issue for research and sustainable development.International Union for the Scientific Study of Population Bulletin, March 2025

    Géraldine Duthé, Arlette Simo Fotso, and Heini Väisänen. The end of the dhs programme: A major issue for research and sustainable development.International Union for the Scientific Study of Population Bulletin, March 2025

    work page 2025

  6. [6]

    Using publicly available satellite imagery and deep learning to understand economic well-being in africa.Nature Communications, 11(1):2583, 2020

    Christopher Yeh, Anthony Perez, Anne Driscoll, George Azzari, Zhongyi Tang, David Lobell, Stefano Ermon, and Marshall Burke. Using publicly available satellite imagery and deep learning to understand economic well-being in africa.Nature Communications, 11(1):2583, 2020

    work page 2020

  7. [7]

    Measuring economic growth from outer space

    J Vernon Henderson, Adam Storeygard, and David N Weil. Measuring economic growth from outer space. American economic review, 102(2):994–1028, 2012

    work page 2012

  8. [8]

    Combining satellite imagery and machine learning to predict poverty.Science, 353(6301):790–794, 2016

    Neal Jean, Marshall Burke, Michael Xie, W Matthew Alampay Davis, David B Lobell, and Stefano Ermon. Combining satellite imagery and machine learning to predict poverty.Science, 353(6301):790–794, 2016

    work page 2016

  9. [9]

    Microestimates of wealth for all low-and middle-income countries.Proceedings of the National Academy of Sciences, 119(3):e2113658119, 2022

    Guanghua Chi, Han Fang, Sourav Chatterjee, and Joshua E Blumenstock. Microestimates of wealth for all low-and middle-income countries.Proceedings of the National Academy of Sciences, 119(3):e2113658119, 2022

    work page 2022

  10. [10]

    Ola Hall, Francis Dompae, Ibrahim Wahab, and Fred Mawunyo Dzanku. A review of machine learning and satellite imagery for poverty prediction: Implications for development research and applications.Journal of International Development, 35(7):1753–1768, 2023

    work page 2023

  11. [11]

    KidSat: satellite imagery to map childhood poverty dataset and benchmark

    Makkunda Sharma, Fan Yang, Duy-Nhat V o, Esra Suel, Swapnil Mishra, Samir Bhatt, Oliver Fiala, William Rudgard, and Seth Flaxman. Kidsat: Satellite imagery to map childhood poverty dataset and benchmark.arXiv preprint arXiv:2407.05986, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Dimensions, indicators, and thresholds for moderate and severe ma- terial shortcomings

    UNICEF. Dimensions, indicators, and thresholds for moderate and severe ma- terial shortcomings. https://data.unicef.org/wp-content/uploads/2019/12/ Child-poverty-moderate-and-severe-material-deprivations-SUMMARY.pdf , 2019. Summary document

    work page 2019

  13. [13]

    Gps data collection

    DHS Program. Gps data collection. https://www.dhsprogram.com/Methodology/GPS-Data.cfm, 2025. Accessed: 2025-11-07

    work page 2025

  14. [14]

    Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy T. V o, Marc Szafraniec, Vladislav Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Gabriel Synnaeve, Hu Xu, Ivan Laptev, Yann LeCun, and Hervé Jegou. Di...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery

    Yezhen Cong, Zongyu Chen, Chenlin Huang, Chuang Zhang, Zheng Zhang, Ming Li, Haoyu Li, and Gong Cheng. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. InAdvances in Neural Information Processing Systems, volume 35, pages 197–211, 2022

    work page 2022

  16. [16]

    A generalizable and accessible approach to machine learning with global satellite imagery

    Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Benjamin Recht, and Solomon Hsiang. A generalizable and accessible approach to machine learning with global satellite imagery. Nature Communications, 12:4392, 2021

    work page 2021

  17. [17]

    Woodcock

    Zhe Zhu and Curtis E. Woodcock. Object-based cloud and cloud shadow detection in landsat imagery.Remote Sensing of Environment, 118:83–94, 2012

    work page 2012

  18. [18]

    Woodcock

    Zhe Zhu, Shixiong Wang, and Curtis E. Woodcock. Improvement and expansion of the fmask algorithm: cloud, cloud shadow, and snow detection for landsats 4–7, 8, and sentinel 2 images.Remote Sensing of Environment, 159:269–277, 2015

    work page 2015

  19. [19]

    Bautista, Jordi Gonzàlez, and Sergio Escalera

    Pau Rodríguez, Miguel A. Bautista, Jordi Gonzàlez, and Sergio Escalera. Beyond one-hot encoding: Lower dimensional target embedding.Image and Vision Computing, 75:21–31, 2018

    work page 2018

  20. [20]

    Similarity encoding for learning with dirty categorical variables

    Patricio Cerda, Gaël Varoquaux, and Balázs Kégl. Similarity encoding for learning with dirty categorical variables. Machine Learning, 107(2):1477–1494, 2018

    work page 2018

  21. [21]

    Encoding high-cardinality string categorical variables

    Patricio Cerdá and Gaël Varoquaux. Encoding high-cardinality string categorical variables.arXiv preprint arXiv:1907.01860, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  22. [22]

    Mathieu J. P. Poirier, Karen A. Grépin, and Michel Grignon. Approaches and alternatives to the wealth index to measure socioeconomic status using survey data: A critical interpretive synthesis.Social Indicators Research, 148:1–46, 2020

    work page 2020

  23. [23]

    Combining disparate data sources for improved poverty prediction and mapping.Proceedings of the National Academy of Sciences, 114(46):E9783–E9792, 2017

    Neeti Pokhriyal and Damien Christophe Jacques. Combining disparate data sources for improved poverty prediction and mapping.Proceedings of the National Academy of Sciences, 114(46):E9783–E9792, 2017

    work page 2017

  24. [24]

    Is poverty predictable with machine learning? a study of dhs data from kyrgyzstan.Socio-Economic Planning Sciences, 81:101195, 2022

    Qing Li, Shuai Yu, Damien Échevin, and Min Fan. Is poverty predictable with machine learning? a study of dhs data from kyrgyzstan.Socio-Economic Planning Sciences, 81:101195, 2022. 9

    work page 2022

  25. [25]

    Pardawala, and Chinmay D

    Kedar Potdar, Taher S. Pardawala, and Chinmay D. Pai. A comparative study of categorical variable encoding techniques for neural network classifiers.International Journal of Computer Applications, 175(4):7–9, October 2017

    work page 2017

  26. [26]

    Accessed: 2025-11-11

    The DHS Program.Standard Recode Manual for DHS-VII (Recode7), September 2018. Accessed: 2025-11-11

    work page 2018

  27. [27]

    Contraceptive failure in the united states.Contraception, 83(5):397–404, May 2011

    James Trussell. Contraceptive failure in the united states.Contraception, 83(5):397–404, May 2011

    work page 2011

  28. [28]

    Spatial analysis of unimproved drinking water source in east africa: Using demographic and health survey (dhs) data from 2012–2023.PLOS ONE, 20(3):e0318189, 2025

    Lidetu Demoze, Kassaw Chekole Adane, Jember Azanaw, Eyob Akalewold, Tenagne Enawugaw, Mitkie Tigabie, Amensisa Hailu Tesfaye, and Gelila Yitageasu. Spatial analysis of unimproved drinking water source in east africa: Using demographic and health survey (dhs) data from 2012–2023.PLOS ONE, 20(3):e0318189, 2025

    work page 2012

  29. [29]

    M. S. Hossain, J. S. Bujang, M. H. Zakaria, and M. Hashim. Assessment of landsat 7 scan line corrector-off data gap-filling methods for seagrass distribution mapping.International Journal of Remote Sensing, 36(4):1188–1215, 2015

    work page 2015

  30. [30]

    Fmask 4.0: Improved cloud and cloud shadow detection in landsats 4–8 and sentinel-2 imagery.Remote Sensing of Environment, 231:111205, 2019

    Shi Qiu, Zhe Zhu, and Binbin He. Fmask 4.0: Improved cloud and cloud shadow detection in landsats 4–8 and sentinel-2 imagery.Remote Sensing of Environment, 231:111205, 2019

    work page 2019

  31. [31]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    work page 2017

  32. [32]

    Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T

    Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    work page 2020

  33. [33]

    Random features for large-scale kernel machines

    Ali Rahimi and Ben Recht. Random features for large-scale kernel machines. InAdvances in Neural Information Processing Systems (NIPS), 2007

    work page 2007

  34. [34]

    Multi-Scale Representation Learning for Spatial Feature Distributions using Grid Cells

    Gengchen Mai, Krzysztof Janowicz, Yingjie Hu, Song Gao, Grant McKenzie, et al. Multi-scale representation learning for spatial feature distributions using grid cells.arXiv preprint arXiv:2003.00824, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  35. [35]

    Sphere2Vec: A General-Purpose Location Representation Learning over a Spherical Surface for Large-Scale Geospatial Predictions

    Gengchen Mai, Yao Xuan, Wenyun Zuo, Yutong He, Jiaming Song, Stefano Ermon, Krzysztof Janowicz, and Ni Lao. Sphere2vec: A general-purpose location representation learning over a spherical surface for large-scale geospatial predictions.arXiv preprint arXiv:2306.17624, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  36. [36]

    Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks

    Marc Rußwurm, Konstantin Klemmer, Esther Rolf, Robin Zbinden, and Devis Tuia. Geographic location encoding with spherical harmonics and sinusoidal representation networks.arXiv preprint arXiv:2310.06743, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  37. [37]

    Random forests.Machine Learning, 45(1):5–32, 2001

    Leo Breiman. Random forests.Machine Learning, 45(1):5–32, 2001

    work page 2001

  38. [38]

    Lightgbm: A highly efficient gradient boosting decision tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. InAdvances in Neural Information Processing Systems (NeurIPS 2017), pages 3146–3154, 2017

    work page 2017

  39. [39]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD 2016), pages 785–794, 2016

    work page 2016

  40. [40]

    A single image deep learning approach to restoration of corrupted landsat-7 satellite images.Sensors, 22(23):9273, 2022

    Anna Petrovskaia, Raghavendra Jana, and Ivan Oseledets. A single image deep learning approach to restoration of corrupted landsat-7 satellite images.Sensors, 22(23):9273, 2022

    work page 2022

  41. [41]

    Deep internal learning for inpainting of cloud-affected regions in satellite imagery.Remote Sensing, 14(6):1342, 2022

    Mikolaj Czerkawski, Priti Upadhyay, Christopher Davison, Astrid Werkmeister, Javier Cardona, Robert Atkinson, Craig Michie, Ivan Andonovic, Malcolm Macdonald, and Christos Tachtatzis. Deep internal learning for inpainting of cloud-affected regions in satellite imagery.Remote Sensing, 14(6):1342, 2022

    work page 2022

  42. [42]

    Deep Image Prior

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior.arXiv preprint arXiv:1711.10925, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  43. [43]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

    work page 2020

  44. [44]

    A hyperspectral image classification approach based on feature fusion and multi-layered gradient boosting decision trees.Entropy, 23(1), 2021

    Shenyuan Xu, Size Liu, Hua Wang, Wenjie Chen, Fan Zhang, and Zhu Xiao. A hyperspectral image classification approach based on feature fusion and multi-layered gradient boosting decision trees.Entropy, 23(1), 2021

    work page 2021

  45. [45]

    Doll, Jan-Peter Muller, and Jeremy G

    Christopher N.H. Doll, Jan-Peter Muller, and Jeremy G. Morley. Mapping regional economic activity from night-time light satellite imagery.Ecological Economics, 57(1):75–92, 2006

    work page 2006

  46. [46]

    Ucdp: The world’s main provider of data on organized violence

    Uppsala Conflict Data Program. Ucdp: The world’s main provider of data on organized violence. https://www. uu.se/en/department/peace-and-conflict-research/research/ucdp/ , 2025. Accessed: 2025-11-12

    work page 2025

  47. [47]

    Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization

    Vicente Vivanco, Gaurav Kumar Nayak, and Mubarak Shah. Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. InAdvances in Neural Information Processing Systems,

    work page

  48. [48]

    arXiv:2309.16020. 10

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    Range: Retrieval augmented neural fields for multi-resolution geo-embeddings, 2025

    Aayush Dhakal, Srikumar Sastry, Subash Khanal, Adeel Ahmad, Eric Xing, and Nathan Jacobs. Range: Retrieval augmented neural fields for multi-resolution geo-embeddings, 2025

    work page 2025

  50. [50]

    Locdiff: Identifying locations on earth by diffusing in the hilbert space, 2025

    Zhangyu Wang, Zeping Liu, Jielu Zhang, Zhongliang Zhou, Qian Cao, Nemin Wu, Lan Mu, Yang Song, Yiqun Xie, Ni Lao, and Gengchen Mai. Locdiff: Identifying locations on earth by diffusing in the hilbert space, 2025. 11 Appendix A DHS Variables Used in the KidSat Setup The table below presents the 17 DHS variables used in the original setup, as well as the ne...

    work page 2025

  51. [51]

    Children aged 12–35 months did not receive immunization against measles nor any dose of DPT

    work page

  52. [52]

    Girls aged 15–17 years have an unmet need for contraception and are not using any contraceptive method

    work page

  53. [53]

    Children experienced an acute respiratory infection (fever or cough) and received no treatment. h3 DPT 1 vaccination h5 DPT 2 vaccination h7 DPT 3 vaccination h9 Measles 1 vaccination h31 Child had cough recently v312 Current contraceptive method Education hv106 Highest education level in house- hold Meeting any one of the following:

    work page

  54. [54]

    Children aged 6–14 who have never been to school

    work page

  55. [55]

    hv109 Educational attainment recoded hv121 School attendance current year New Indicators hv025 Rural/Urban Status New indicators

    Children aged 15–17 who have not com- pleted primary school. hv109 Educational attainment recoded hv121 School attendance current year New Indicators hv025 Rural/Urban Status New indicators. hv270 Wealth index 12 B Re-aggregation of One-Hot Encoding Columns Note: In DHS surveys, each indicator corresponds to the response to a particular question, and each...

    work page 2000