Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Joint Utilization of Geospatial and census proxies for Autoencoder-Assisted Downscaling (JUGAAD) of socioeconomic indicators in India

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that socioeconomic indicators collected only at district level can be reconstructed at village-cluster scale by training an autoencoder-regression pipeline on census, geospatial, and state features, validated by aggregating

desk verdict Useful pipeline with honest limitations, but the headline accuracy claim rests on in-sample evaluation; needs a real out-of-sample test before it can be believed. read the letter →

arxiv 2607.20559 v1 pith:HMQD23P2 submitted 2026-07-17 cs.LG cs.AI

Aditya Dutt , Paul Gader , Aditya Singh This is my paper
classification cs.LGcs.AI
keywords autoencodersspatialdownscalingsocioeconomicindicatorspovertymappingfoodsecuritycensusdatamachinelearningIndia
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

JuGAAD claims that the detailed socioeconomic indicators India's National Sample Survey Office collects at district level can be generated at village-cluster scale from public census, geospatial, and state-level features alone during inference. The authors build a three-stage pipeline: a hexagonal tessellation of about 20 villages provides consistent spatial units; per-category autoencoders compress 475 NSSO features into a 112-dimensional latent space; and a regression network learns to map 141 census, geospatial, and state features to that latent space at district level, then applies the mapping to hexagons. When these fine-scale predictions are aggregated back to districts, they match ground-truth NSSO indicators well, especially for consumer expenditure. A sympathetic reader would care because it offers a path to sub-district poverty and food-security maps for policy in data-scarce contexts, and the maps reproduce known regional patterns like coffee consumption in southern India and tube-well reliance in the north.

What carries the argument

The load-bearing mechanism is the combination of a hexagonal village-cluster tessellation with an autoencoder-regression-decoder stack. Each hexagon (about 20 villages, 15 km vertex-to-vertex) regularizes administrative boundary changes and provides a stable inference unit; six category-specific autoencoders compress 475 collinear NSSO indicators to 112 latent dimensions; a fully connected regression network maps a 141-dimensional input (58 census, 48 geospatial, and 35 state identifiers) to these latent dimensions; and the decoders reconstruct the full indicators. This stack lets the model train where labeled data exist (districts) and predict where they do not (hexagons), which is the enti

What would settle it

Take districts that have additional village- or taluk-level survey data (or a recent disaggregated sample), run the trained model to produce hexagon-level predictions for those villages, and compare directly against the fine-grained ground truth. If the hexagon-level predictions diverge systematically even though their district averages match, the central downscaling claim is an artifact of aggregation. A secondary check: train the regression on 2011 district data and predict 2001; since the autoencoder is trained on combined years, cross-year transfer tests whether the census-to-latent mappin

Watch

Extended reading notes

Core claim

The paper's central claim is that the relationship learned at district level between census, geospatial, and state features and the compressed latent representation of NSSO survey data can be applied at hexagon level to produce meaningful high-resolution estimates, without any survey data at inference. The authors validate this by aggregating the hexagon-level predictions back to districts and comparing against ground truth, reporting strong accuracy across most categories and confirming the central premise of the approach. The decoded maps reproduce broad spatial patterns (coffee expenditure concentrated in southern states, inverse tap water and tube well patterns, pucca housing prevalence)

Load-bearing premise

The district-level relationship between census, geospatial, and state features and the NSSO latent encodings holds at the hexagon and village-cluster level, with no distributional shift between the features used in training and those used at inference.

Editorial extensions

If this is right

  • High-resolution maps of all 475 NSSO indicators become available at village-cluster scale for 2001 and 2011, enabling sub-district poverty and food-security analysis across all of India.
  • Because only census, geospatial, and state features are needed at inference, the pipeline can be rerun for other years or regions that have similar public data, without new surveys.
  • The downscaled maps reveal plausible intra-district heterogeneity and temporal shifts (e.g., coffee consumption spreading beyond the south, expansion of tap water access, decline in meal-skipping), giving policy-relevant texture absent from district data.
  • The approach outperforms PCA for reducing the dimensionality of the survey data, supporting the use of nonlinear autoencoders for collinear socioeconomic features.
  • End-to-end fidelity is category-dependent: consumer expenditure and land/livestock indicators reconstruct well, while agriculture and employment remain weak, meaning the claim of strong accuracy properly applies to a subset of indicators.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the district-to-hexagon transfer generalizes, the same framework could downscale other coarse surveys, such as NFHS, or combine multiple surveys into one latent space, producing a richer set of sub-district indicators.
  • A direct test the authors do not report is cross-year transfer: training on 2011 and predicting 2001 (or vice versa) would probe whether the census-to-latent mapping is temporally stable or year-specific.
  • The hexagon tessellation is agnostic to Indian administrative boundaries, so the approach could extend to other countries with census and survey data, provided a comparable geospatial feature set exists.
  • The apparent accuracy could be partly driven by district-level aggregation smoothing away hexagon-level errors; comparing against a handful of village-level ground-truth surveys would settle whether the fine-scale patterns are real.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes JuGAAD, a three-stage pipeline for downscaling district-level NSSO socioeconomic indicators to a hexagonal village-cluster grid in India. First, an autoencoder compresses 475 NSSO features into 112 latent dimensions. Second, a fully connected regression network maps district-level census, geospatial, and state features to those latent encodings. Third, the regression is applied to hexagon-level census/geospatial features and the resulting latent vectors are decoded into high-resolution NSSO-indicator predictions. The authors report reconstruction R² values, district-level regression R² values, and end-to-end R² values after aggregating cluster predictions back to districts, and they present qualitative maps. The central claim is that the method 'predicts socioeconomic indicators at fine scales with strong accuracy.'

Significance. If the central claim were established, JuGAAD would be a useful addition to the small-area-estimation and poverty-mapping literature: it combines a nonlinear autoencoder with a hexagon-based spatial framework and uses only public census/geospatial inputs at inference. The idea of compressing high-dimensional survey indicators and then regressing on coarse features is a reasonable strategy, and the topic is of clear policy relevance. However, the paper's headline result is not supported by its evaluation design. All quantitative assessments are effectively in-sample: the autoencoder is trained and evaluated on the same district-level NSSO data, the regression is trained on district-level features and evaluated on the same districts, and the end-to-end metrics aggregate cluster predictions to those same districts. No held-out split, spatial cross-validation, temporal holdout, or independent cluster-level benchmark is provided. Consequently, the paper's main claim of fine-scale accuracy is unverified, and the reported R² values are consistent with interpolation of the training districts rather than generalization. The paper also explicitly acknowledges a distributional shift between d

major comments (4)
  1. [§2.1, Tables 2–4, §3.2.5] The evaluation is completely in-sample. Table 2 is explicitly labeled 'during training'; Tables 3 and 4 aggregate cluster predictions back to the district level and compare them to the same district-level NSSO encodings and indicators used to train the autoencoder (§3.2.1) and the regression (§3.2.2). There is no train/test split, cross-validation, spatial holdout, or temporal holdout. Training R² ≈ 0.99 (Table 2) shows only that the regression can fit its inputs; the lower but positive R² values in Tables 3–4 can be produced by a model that interpolates the training districts and does not generalize to unseen areas. The abstract's claim of 'strong accuracy' at fine scales is therefore not supported by the reported evidence.
  2. [§3.2.5, §2.1] District-level aggregation cannot validate fine-scale predictions. In the end-to-end evaluation, cluster-level predictions are averaged to the district scale before computing MSE/R² (§3.2.5, Table 4). A model that predicted the district-level mean for every hexagon would achieve the same aggregated R² while containing no real sub-district information. The claimed output is high-resolution village-cluster estimates, but the only quantitative comparison is at the district level. Unless the evaluation demonstrates that predicted cluster-level values carry information beyond the district mean (e.g., by comparing against sub-district survey data, synthetic cluster targets, or at least a district-mean baseline), the fine-scale claim is not tested.
  3. [§4 Limitations; §3.2.2–§3.2.4] The distributional-shift limitation is load-bearing and untreated. The regression REG(·) is trained on district-level inputs X_dist_inp (Eq. 8) and then applied to hexagon-level inputs X_hex_inp (Eq. 11). The decoder is trained exclusively on latent vectors from district-level NSSO data but decodes latent vectors produced from hexagon-level features. The authors acknowledge that 'a potential distributional shift may exist between training and inference phases' and that 'the statistical properties of these two sets of latent vectors may differ.' This is not a minor caveat: it directly threatens the central premise that the learned mapping transfers to the fine scale. The paper provides no diagnostic, such as a comparison of district-level versus hexagon-level feature distributions, a synthetic-cluster experiment, or a spatial holdout that would test transfer. Without such evidence, the fi
  4. [Table 1, Fig. 1] The autoencoder evaluation is also performed on training data. Table 1 reports reconstruction R² from the combined 2001/2011 NSSO dataset, and Fig. 1's caption states 'on training data.' Reconstruction fidelity on the training set does not establish that the latent representation is informative for unseen districts or that the 112-dimensional bottleneck preserves all downstream-relevant information. A held-out reconstruction evaluation (e.g., by district or by year) is needed before the latent space can be treated as a reliable target representation for regression.
minor comments (5)
  1. [§3.1] Typographical and grammatical errors: 'This methodology is designed to leleveragesgths of each dataset' appears garbled; please revise.
  2. [§2.1, Table 2] The heading of Table 2 says 'during training,' but the text in §2.1 presents it as 'predictive performance.' Clarify that these are in-sample fits and do not represent generalization.
  3. [§3.1.1, Appendix A] The imputation acceptance threshold is stated as R² ≥ 0.6, but Appendix A reports mean R² values above 0.94. Clarify whether the threshold is for acceptance after cross-validation or for a different stage, and report how many features were removed at each step.
  4. [§4, Data Availability] The paper states that cleaned data and predictions 'will be made publicly available,' but no repository link or code is provided. For reproducibility, please include URLs or DOIs in the final manuscript.
  5. [References] Some references contain incomplete bibliographic information (e.g., [1], [28], [38]) and URLs that are not consistently formatted. Please standardize.

Circularity Check

1 steps flagged · score 6.0 of 10

Validation of the fine-scale claim compares aggregated cluster predictions to the same district-level NSSO encodings used to train the regression; no held-out split or cluster-level ground truth exists, so the headline accuracy is an in-sample fit.

  1. fitted input called prediction [§2.1 (quantitative evaluation); §3.2.3; §3.2.5]
    "For the regression model training phase, y=z^{dist,x}, and \tilde{y}=\tilde{z}^{dist,x}. ... To validate these downscaled predictions, the predicted cluster-level encodings were then aggregated to the district level, enabling a direct comparison with the ground-truth district-level encodings."

    REG(·) is trained to minimize error against z^{dist,x}, the district-level NSSO latent encodings (§3.2.3). The validation pipeline takes cluster predictions \tilde{z}^{hex,x}=REG(X^{hex}_{inp}), averages them to districts, and compares them with those same z^{dist,x} encodings (§2.1, §3.2.5). No held-out district, year, or independent cluster-level ground truth is described, so Tables 3–4 report in-sample fit, not out-of-sample prediction. The abstract's claim that predictions are confirmed 'with strong accuracy' and §2.1's statement that 'this validates the central premise' therefore reduce to the fitted regression's training-target comparison, not to a test of the fine-scale transfer.

full rationale

The central claimed derivation—census/geospatial features → NSSO indicators—is not definitionally circular: the target encodings come from NSSO, and the cluster predictions are not algebraically identical to the training outputs. However, the only quantitative validation reported (Tables 3–4) uses the same district-level NSSO data that defined the regression targets. In §3.2.3 the regression is trained with y = z^{dist,x}; §3.2.5 validates by aggregating cluster predictions to districts and computing MSE against 'actual district-level NSSO indicators.' No held-out districts, temporal holdout, cross-validation, or independent cluster-level survey is described anywhere. Thus the abstract's 'predicts ... at fine scales with strong accuracy' is supported only by in-sample fit, and the paper's own limitations acknowledge a single sample per district and a training/inference distributional shift (§4). The 'fine-scale' component is especially unvalidated: district-level aggregation cannot distinguish genuine sub-district structure from a model that reproduces district means. No load-bearing self-citation chain appears in the references. The circularity is therefore of the fitted-input-called-prediction form: the fitted model's performance on its own training targets is renamed as validation/prediction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The method rests on the transferability of a district-learned mapping to hexagons, on the reliability of imputed NSSO values, and on hand-picked tessellation and latent-dimensionality choices. None of these is independently tested in the preprint.

free parameters (5)
  • Hexagon size = 15 km vertex-to-vertex
    Chosen by the authors to balance noise reduction and spatial resolution; no sensitivity analysis is reported (§3.1).
  • Minimum villages per hexagon = 5
    Hexagons with fewer than 5 villages dropped to avoid sparse extrapolation (§3.1).
  • Latent dimensionality per category = 16–20 per category, 112 total
    Hand-selected latent dimensions; no variation or reconstruction-based selection is reported (Table 5, §3.2.1).
  • Imputation acceptance threshold = R² ≥ 0.6
    Features that could not be imputed above this threshold across predictor combinations were removed, affecting the 475-feature set (§3.1.1).
  • Neural network hyperparameters = not specified
    Hidden widths, number of blocks, Huber δ, learning rate, epochs, and ensemble size are not reported; they are chosen by hand (§3.2.2–3.2.3).
assumptions (5)
  • domain assumption District-level NSSO data are an unbiased ground truth after imputation; imputed values do not distort downstream evaluation.
    §3.1.1 and Appendix A: features with missing values are imputed using shallow neural networks; features failing R²≥0.6 are dropped. If imputation errors correlate with census predictors, the regression inherits them.
  • domain assumption The district-level relationship between census/geospatial/state features and NSSO latent encodings transfers unchanged to hexagon-level inputs.
    §3.2.4 inference uses REG(·) trained on district inputs at hexagon scale; §4 Limitations admits a potential distributional shift between training and inference phases. This is the load-bearing transfer assumption.
  • domain assumption Hexagonal tessellation with majority-rule assignment is a valid, noise-reducing spatial unit that preserves district/subdistrict structure.
    §3.1: size and village threshold are chosen by the authors; no sensitivity analysis is provided.
  • domain assumption The 141 input features (58 census + 48 geospatial + 35 state dummies) contain enough signal to predict the 112-d NSSO latent space at cluster level.
    §3.1.2 and §3.2.2; this is the modeling premise of the regression stage.
  • ad hoc to paper Autoencoder latent dimensions preserve all information relevant to downstream prediction; reconstruction R²≈0.85–0.89 is sufficient.
    §3.2.1/Table 1: latent dimensionalities are chosen per category and not varied; reconstruction is measured on training data only.
invented entities (2)
  • Hexagonal village-cluster tessellation
    purpose: Stable spatial unit for fine-scale downscaling and cross-year alignment
    A constructed spatial framework with no independent validation; predictions and interpretation depend on this unit.
  • NSSO latent representation (112-d)
    purpose: Compressed target for the regression model and for decoding high-resolution indicators
    An internal learned construct; it has no direct external handle and is evaluated only through reconstruction and downstream metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Joint Utilization of Geospatial and census proxies for Autoencoder-Assisted Downscaling (JUGAAD) of socioeconomic indicators in India." pith.science (2026). https://pith.science/paper/HMQD23P2

@misc{pith2026260720559,
  author       = {Pith},
  title        = {Pith review of: Joint Utilization of Geospatial and census proxies for Autoencoder-Assisted Downscaling (JUGAAD) of socioeconomic indicators in India},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HMQD23P2}},
  note         = {Machine review of arXiv:2607.20559}
}
read the original abstract

Monitoring poverty and food security indicators is imperative for addressing socioeconomic challenges in developing nations. A limitation is mismatches in scale between data sources: census data provide geographic coverage, while socioeconomic indicators are derived from infrequently conducted surveys at coarse resolutions, posing a methodological challenge. This study introduces a deep learning framework, JuGAAD, using Indian census and survey data from 2001 and 2011 as a case study. We employ a three-step process: census and geospatial data are averaged into intermediate village-cluster-scale tessellations to reduce noise and regularize administrative boundary changes; an autoencoder compresses high-dimensional National Sample Survey Office (NSSO) data into a low-dimensional latent representation; and a regression model maps upscaled census and geospatial data to this representation. This function is applied to fine-grained census data to generate high-resolution predictions, validated against ground-truth district-level NSSO indicators. Results confirm the methodology predicts socioeconomic indicators at fine scales with strong accuracy.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 4 canonical work pages

  1. [1]

    ISI26(31) (2015)

    Das, S., Chambers, R.: A robust ell methodology for poverty mapping. ISI26(31) (2015)

    2015

  2. [2]

    Canadian Journal of Statistics38(3), 369–385 (2010) https://doi.org/10.1002/cjs.10051

    Molina, I., Rao, J.N.K.: Small area estimation of poverty indicators. Canadian Journal of Statistics38(3), 369–385 (2010) https://doi.org/10.1002/cjs.10051 . eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/cjs.10051. Accessed 2021-10-24

    work page doi:10.1002/cjs.10051 2010

  3. [3]

    Biometrika93, 255–268 (2006) https://doi.org/10.1093/biomet/93.2.255

    Chambers, R., Tzavidis, N.: M-Quantile Models for Small Area Estimation. Biometrika93, 255–268 (2006) https://doi.org/10.1093/biomet/93.2.255

    work page doi:10.1093/biomet/93.2.255 2006

  4. [4]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 1885–1894 (2021)

    Huang, X., Shi, K., Cui, Y., Li, Y.: A saturated light correction method for dmsp-ols nighttime stable light data by remote and social sensing data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14, 1885–1894 (2021)

    2021

  5. [5]

    Mellander, C., Lobo, J., Stolarick, K., Matheson, Z.: Night-time light data: A good proxy measure for economic activity? PloS one10(10), 0139779 (2015)

    2015

  6. [6]

    Population Health Metrics6(1), 5 (2008)

    Noor, A.M., Alegana, V.A., Gething, P.W., Tatem, A.J., Snow, R.W.: Using remotely sensed night-time light as a proxy for poverty in africa. Population Health Metrics6(1), 5 (2008)

    2008

  7. [7]

    ISPRS International Journal of Geo-Information8(11), 498 (2019)

    Andersson, M., Hall, O., Archila, M.F.: How data-poor countries remain data poor: Underestimation of human settlements in burkina faso as observed from nighttime light data. ISPRS International Journal of Geo-Information8(11), 498 (2019)

    2019

  8. [8]

    Science 353(6301), 790–794 (2016)

    Jean, N., Burke, M., Xie, M., Alampay Davis, W.M., Lobell, D.B., Ermon, S.: Combining satellite imagery and machine learning to predict poverty. Science 353(6301), 790–794 (2016)

    2016

Show all 42 references
  1. [9]

    Proceedings of the IEEE86(11), 2278–2324 (2002)

    LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE86(11), 2278–2324 (2002)

    2002

  2. [10]

    Social Indicators Research167(1), 475–505 (2023)

    Daoud, A., Jord´ an, F., Sharma, M., Johansson, F., Dubhashi, D., Paul, S., Baner- jee, S.: Using satellite images and deep learning to measure health and living standards in india. Social Indicators Research167(1), 475–505 (2023)

    2023

  3. [11]

    arXiv preprint arXiv:1312.6229 (2013) 34

    Sermanet, P., Eigen, D., Zhang, X., Mathieu, M., Fergus, R., LeCun, Y.: Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint arXiv:1312.6229 (2013) 34

    2013 arXiv

  4. [12]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp

    Sharif Razavian, A., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off- the-shelf: an astounding baseline for recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 806– 813 (2014)

    2014

  5. [13]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Xie, M., Jean, N., Burke, M., Lobell, D., Ermon, S.: Transfer learning from deep features for remote sensing and poverty mapping. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 30 (2016)

    2016

  6. [14]

    In: IJCAI, pp

    Pettersson, M.B., Kakooei, M., Ortheden, J., Johansson, F.D., Daoud, A.: Time series of satellite imagery improve deep learning estimates of neighborhood-level poverty in africa. In: IJCAI, pp. 6165–6173 (2023)

    2023

  7. [15]

    IEEE Transactions on Geoscience and Remote Sensing62, 1–13 (2024)

    Kakooei, M., Daoud, A.: Increasing the confidence of predictive uncertainty: earth observations and deep learning for poverty estimation. IEEE Transactions on Geoscience and Remote Sensing62, 1–13 (2024)

    2024

  8. [16]

    In: 2015 Joint Urban Remote Sensing Event (JURSE), pp

    Engstrom, R., Sandborn, A., Yu, Q., Burgdorfer, J., Stow, D., Weeks, J., Graesser, J.: Mapping slums using spatial features in accra, ghana. In: 2015 Joint Urban Remote Sensing Event (JURSE), pp. 1–4 (2015). IEEE

    2015

  9. [17]

    SSRN Scholarly Paper ID 3785116, Social Science Research Network, Rochester, NY (December 2020)

    Puttanapong, N., Martinez, J., Addawe, M., Bulan, J., Durante, R.L., Mar- tillan, M.: Predicting Poverty Using Geospatial Data in Thailand. SSRN Scholarly Paper ID 3785116, Social Science Research Network, Rochester, NY (December 2020). https://doi.org/10.2139/ssrn.3785116 . h...

    2020 doi

  10. [18]

    Sustainability13(16), 8717 (2021) https://doi.org/10.3390/su13168717

    Xu, Y., Mo, Y., Zhu, S.: Poverty Mapping in the Dian-Gui-Qian Contiguous Extremely Poor Area of Southwest China Based on Multi-Source Geospatial Data. Sustainability13(16), 8717 (2021) https://doi.org/10.3390/su13168717 . Number: 16 Publisher: Multidisciplinary Digital Publish...

    2021 doi

  11. [19]

    Tingzon, I., Orden, A., Go, K.T., Sy, S., Sekara, V., Weber, I., Fatehkia, M., Garc´ ıa-Herranz, M., Kim, D.: MAPPING POVERTY IN THE PHILIPPINES USING MACHINE LEARNING, SATELLITE IMAGERY, AND CROWD-SOURCED GEOSPATIAL INFORMATION. The Interna- tional Archives of the Photogramme...

    2019

  12. [20]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Ayush, K., Uzkent, B., Tanmay, K., Burke, M., Lobell, D., Ermon, S.: Efficient poverty mapping from high resolution remote sensing images. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 12–20 (2021)

    2021

  13. [21]

    35 In: Thirtieth AAAI Conference on Artificial Intelligence (2016)

    Xie, M., Jean, N., Burke, M., Lobell, D., Ermon, S.: Transfer Learn- ing from Deep Features for Remote Sensing and Poverty Mapping. 35 In: Thirtieth AAAI Conference on Artificial Intelligence (2016). https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/view/12196Accessed 2021-10-25

    2016

  14. [22]

    Nature Machine Intelligence1(5), 206–215 (2019) https://doi.org/10.1038/s42256-019-0048-x

    Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence1(5), 206–215 (2019) https://doi.org/10.1038/s42256-019-0048-x . Bandiera abtest: a Cg type: Nature Research Journals Number:...

    2019 doi

  15. [23]

    IEEE Transactions on Emerging Topics in Computational Intelligence 5(5), 726–742 (2021) https://doi.org/10.1109/TETCI.2021.3100641

    Zhang, Y., Tiˇ no, P., Leonardis, A., Tang, K.: A Survey on Neural Network Inter- pretability. IEEE Transactions on Emerging Topics in Computational Intelligence 5(5), 726–742 (2021) https://doi.org/10.1109/TETCI.2021.3100641 . Conference Name: IEEE Transactions on Emerging To...

    2021

  16. [24]

    arXiv:2002.01612 [cs] (2020)

    Ayush, K., Uzkent, B., Burke, M., Lobell, D., Ermon, S.: Generating Interpretable Poverty Maps using Object Detection in Satellite Images. arXiv:2002.01612 [cs] (2020). arXiv: 2002.01612. Accessed 2021-10-25

    2002 arXiv

  17. [25]

    Nature Machine Intelligence2(11), 684–692 (2020)

    Abitbol, J.L., Karsai, M.: Interpretable socioeconomic status inference from aerial imagery through urban patterns. Nature Machine Intelligence2(11), 684–692 (2020)

    2020

  18. [26]

    arXiv preprint arXiv:2011.13563 (2020)

    Ledesma, C., Garonita, O.L., Flores, L.J., Tingzon, I., Dalisay, D.: Inter- pretable poverty mapping using social media data, satellite images, and geospatial information. arXiv preprint arXiv:2011.13563 (2020)

    2011 arXiv

  19. [27]

    Oxford University Press, USA, ??? (2015)

    Alkire, S., Roche, J.M., Ballon, P., Foster, J., Santos, M.E., Seth, S.: Multidi- mensional Poverty Measurement and Analysis. Oxford University Press, USA, ??? (2015)

    2015

  20. [28]

    Quality & Quantity58(4), 3171–3213 (2024)

    D’Attoma, I., Matteucci, M.: Multidimensional poverty: an analysis of definitions, measurement tools, applications and their evolution over time through a system- atic review of the literature up to 2019. Quality & Quantity58(4), 3171–3213 (2024)

    2019

  21. [29]

    The World Bank Economic Review36(3), 605–628 (2022)

    Decerf, B., Ferrando, M.: Unambiguous trends combining absolute and relative income poverty: new results and global application. The World Bank Economic Review36(3), 605–628 (2022)

    2022

  22. [30]

    Journal of Poverty24(3), 241–255 (2020)

    Chung, Y., Maguire-Jack, K.: Understanding movement into poverty and poverty persistence over time. Journal of Poverty24(3), 241–255 (2020)

    2020

  23. [31]

    The Journal of Development Studies55(7), 1527–1547 (2019)

    Dang, H.-A.H., Dabalen, A.L.: Is poverty in africa mostly chronic or transient? 36 evidence from synthetic panel data. The Journal of Development Studies55(7), 1527–1547 (2019)

    2019

  24. [32]

    domains of deprivation framework

    Abascal, A., Rothwell, N., Shonowo, A., Thomson, D.R., Elias, P., Elsey, H., Yeboah, G., Kuffer, M.: “domains of deprivation framework” for mapping slums, informal settlements, and other deprived areas in lmics to improve urban planning and policy: A scoping review. Computers,...

    2022

  25. [33]

    Patterns3(10) (2022)

    Hall, O., Ohlsson, M., R¨ ognvaldsson, T.: A review of explainable ai in the satellite data, deep machine learning, and human poverty domain. Patterns3(10) (2022)

    2022

  26. [34]

    arXiv preprint arXiv:2202.00109 (2021)

    Daoud, A., Jordan, F., Sharma, M., Johansson, F., Dubhashi, D., Paul, S., Baner- jee, S.: Measuring poverty in india with machine learning and remote sensing. arXiv preprint arXiv:2202.00109 (2021)

    2021 arXiv

  27. [35]

    In: Predicting Inequality of Opportunity and Poverty in India Using Machine Learning, pp

    Mehta, B.S., Srivastava, R., Dhote, S.: Predicting poverty with machine learning and geospatial data. In: Predicting Inequality of Opportunity and Poverty in India Using Machine Learning, pp. 75–109. Springer, ??? (2025)

    2025

  28. [36]

    Social Indicators Research, 1–24 (2025)

    Arya, P.K., Sur, K., Dhote, S., Siral, H., Kundu, T., Mehta, B.S., Srivastava, R.: Integrating multi-source satellite imagery and socio-economic household data for wealth-based poverty assessment of india: A gis and machine learning based approach: Arya et al. Social Indicator...

    2025

  29. [37]

    In: 2024 IEEE International Students’ Conference on Electrical, Electronics and Computer Science (SCEECS), pp

    Khare, P., Jadhav, D.: Eyes on poverty: A comprehensive review of computer vision and textual data integration for poverty prediction in india. In: 2024 IEEE International Students’ Conference on Electrical, Electronics and Computer Science (SCEECS), pp. 1–7 (2024). IEEE

    2024

  30. [38]

    Agricultural economics research review 31(2), 231–240 (2018)

    Subash, S.P., Kumar, R.R., Aditya, K.S.: Satellite data and machine learning tools for predicting poverty in rural india. Agricultural economics research review 31(2), 231–240 (2018)

    2018

  31. [39]

    Chemometrics and Intelligent Laboratory Systems2(1), 37–52 (1987) https://doi.org/10.1016/ 0169-7439(87)80084-9

    Wold, S., Esbensen, K., Geladi, P.: Principal component analysis. Chemometrics and Intelligent Laboratory Systems2(1), 37–52 (1987) https://doi.org/10.1016/ 0169-7439(87)80084-9 . Accessed 2021-10-25

    1987

  32. [40]

    In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery And Data Mining

    Zhou, C., Paffenroth, R.C.: Anomaly Detection with Robust Deep Autoen- coders. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery And Data Mining. KDD ’17, pp. 665–674. Association for Computing Machinery, New York, NY, USA (2017). https://d...

    2017

  33. [41]

    arXiv:1603.00982 [cs] (2016)

    Chung, Y.-A., Wu, C.-C., Shen, C.-H., Lee, H.-Y., Lee, L.-S.: Audio Word2Vec: Unsupervised Learning of Audio Segment Representations using Sequence-to- sequence Autoencoder. arXiv:1603.00982 [cs] (2016). arXiv: 1603.00982. Accessed 37 2021-10-25

    2016 arXiv

  34. [42]

    In: Advances in Neural Information Pro- cessing Systems, vol

    Pu, Y., Gan, Z., Henao, R., Yuan, X., Li, C., Stevens, A., Carin, L.: Variational Autoencoder for Deep Learning of Images, Labels and Captions. In: Advances in Neural Information Pro- cessing Systems, vol. 29. Curran Associates, Inc., ??? (2016). https://proceedings.neurips.cc...

    2016

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.