Pith. sign in

REVIEW 4 major objections 4 minor 63 references

Open-source vision-language models, prompted on street-view photos, can estimate residential blight citywide, and stacking three models' outputs into an XGBoost ensemble yields more accurate and robust assessments than any single model.

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 ↗

Open-source vision-language models with an XGBoost ensemble can assess residential blight from street-view images, but reported performance is inflated by test-set threshold tuning.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A useful applied paper whose central ensemble-advantage claim is undermined by test-set threshold tuning and an internal contradiction. the 4 major comments →

arxiv 2608.01753 v1 pith:RT5YUIRH submitted 2026-08-03 cs.CV cs.AI

Can Urban Blight Be Accessed with Vision-language Models: A Case Study in Detroit

Xiaohao Yang , Aohua Tian , Derek Van Berkel , Xu Qiang , Mark Lindquist This is my paper
classification cs.CV cs.AI
keywords urban blightvision-language modelsstreet view imageryensemble stackingXGBoostresidential housing conditionDetroitmultiple views
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.

The reading

Traditional blight surveys require trained assessors walking or driving every block, which is slow and expensive to repeat. This paper tries to establish that a cheaper pipeline—open-source vision-language models looking at three street-view photos per house and answering structured questions about roof, window/door, and facade damage—can produce citywide residential-blight estimates that track professional in-person ratings. The central result is that no single model is best on everything: Gemma3 finds broken windows well but overpredicts, Qwen2.5-VL is precise but misses damage, and Mistral-small3.1 sits in between. Feeding all three models' probabilities and pairwise agreements into an XGBoost stacked ensemble gives higher accuracy, F1, and Cohen's Kappa than any base model across all three damage categories. If this holds, cities can update housing-condition maps at low cost between formal surveys, using the probability outputs to tune how aggressively they screen.

Core claim

The paper's central claim is that residential blight can be assessed from street-level imagery by open-source vision-language models when each house is viewed from multiple perspectives and the models are asked, via structured prompts, for both a binary judgment and a calibrated probability of damage for roof, openings, and facade. Evaluated against 18,886 Detroit Land Bank Authority professional in-person assessments in Detroit, the three VLMs show complementary inference strengths, and an XGBoost meta-learner trained on their outputs—including probability summaries and agreement/difference features—outperforms every individual model on accuracy, F1, and Kappa. The ensemble also has the low

What carries the argument

The load-bearing mechanism is stacked generalization: three open-source VLMs (Gemma3 27B, Mistral-small3.1 24B, Qwen2.5-VL 32B), quantized to 4 bits and run at temperature 0, independently score each of three street views per property on roof, window/door, and facade damage. Their outputs are condensed into 18 engineered features—per-model binary vote, vote count, mean/min/max/SD probability, and pairwise agreement and difference—which an XGBoost meta-learner combines. A blight score weights damages by severity (+3 roof, +1.5 opening, +4 facade). The decision threshold is chosen on the test set to maximize Cohen's Kappa or F1, and the threshold objective controls the precision–recall trade-o

Load-bearing premise

The evaluation tunes the decision threshold on the held-out test set to maximize Cohen's Kappa; the central claim assumes this is an honest measure of generalization, rather than fixing thresholds before seeing the test set.

What would settle it

Recompute the ensemble-versus-best-base comparison with the threshold fixed on a validation split, or via nested cross-validation, before any test-set evaluation. If the ensemble's Kappa or accuracy advantage over the best single model disappears, the claim that the ensemble learner outperforms individual base models is not supported by this data.

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

If this is right

  • Multiple street views of the same property reliably improve kappa, accuracy, recall, and F1 over any single view, at the cost of some precision and specificity.
  • Different VLMs are complementary: Gemma3 leads on openings but overpredicts, Qwen2.5-VL is conservative with high precision and low recall, and Mistral-small3.1 is balanced.
  • The stacked ensemble beats every base model on accuracy, F1, Kappa, RMSE, MAE, and concordance, with the largest gains on facade and window/door conditions.
  • Practitioners can tune the threshold to prioritize recall for broad screening or precision for enforcement, and use probability surfaces to identify high-risk hotspots.
  • The framework provides a low-cost, regularly updatable complement to formal blight surveys, guiding field-inspection triage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because thresholds are tuned on the held-out test set to maximize Kappa, the reported ensemble gains likely overstate what a deployment would see; fixing thresholds on a validation split before test evaluation is the natural follow-up re-analysis.
  • The SHAP result that mean predicted probability dominates feature importance suggests a simple average of model probabilities may capture much of the ensemble's benefit; the paper's more elaborate agreement features contribute less.
  • The framework should transfer to other cities and to storefront vacancy, yard maintenance, or post-disaster damage screening, but transferability is untested here—Detroit's housing styles, image distribution, and blight base rates may differ.
  • A testable extension is to check whether increasing ensemble diversity, through more base VLMs or varied temperature settings, improves robustness without fine-tuning, which the authors list as future work.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 4 minor

Summary. The paper proposes a scalable framework for estimating residential blight in Detroit using three open-source vision-language models (Gemma 3, Mistral-small 3.1, Qwen2.5-VL) applied to multiple street-view images per property. Structured prompts elicit binary damage judgments and probabilistic estimates for roof, window/door, and facade conditions, which are compared with professional Detroit Land Bank Authority annotations. A weighted blight score is derived, and an XGBoost stacked ensemble combines the base-model outputs with engineered features. The authors report that multiple street views help, that the three models have complementary strengths, and that the ensemble learner outperforms all individual models. They position the framework as a low-cost, regularly updatable complement to in-person surveys.

Significance. The application is timely and practically important: if validated, the framework could give municipalities a low-cost, scalable, and regularly updatable complement to labor-intensive blight surveys. The evaluation against independent professional annotations is a genuine strength and is not circular. The use of open-source models and a stacking approach is also a sensible design choice. However, the paper's central empirical claim—ensemble superiority over all base models—is not supported by the evaluation as reported because the decision threshold is tuned on the test set, the comparison with base models is asymmetric, and the paper's own numbers contain a direct contradiction with the abstract. The practical relevance is high, but the current evidence base needs substantial revision before the claims can be accepted.

major comments (4)
  1. [§3.3 and §4.3] The decision threshold for the ensemble is selected by sweeping 0.10–0.90 and choosing the threshold that maximizes Cohen's Kappa, and then the reported metrics are 'based on the test set' (§4.3). No separate validation split is described. This makes the threshold a test-set-fitted parameter, so the reported Kappa/F1/accuracy values are in-sample estimates for that threshold choice. If base models are evaluated with their native binary decision rule and did not receive the same threshold sweep, the comparison is also asymmetric. A clean protocol would fix thresholds on a training/validation split (or via nested cross-validation) before any test-set evaluation. As presented, the headline claim that the ensemble learner outperforms all base models is not supported.
  2. [§4.3, Fig. 6, Abstract, Conclusion] The text states that the ensemble 'attains higher accuracy (0.7–0.77), F1-score (0.71–0.84), and kappa (0.41–0.43) across all residential conditions compared to the base models, with a slight drop in accuracy and F1-score compared to Gemma3.' Gemma3 is one of the base models, so the unqualified statements in the abstract and conclusion ('the ensemble learner outperforms individual base models', 'outperforming each base model') are contradicted by the paper's own reported numbers. The authors should re-specify the primary metric and provide a statistical comparison; if the claim is only 'better on Kappa and error measures, but not on all metrics', the abstract and conclusion must be revised accordingly.
  3. [§4.1, Fig. 3, Appendix C] The claimed benefit of multiple street views is partly an artifact of the OR-aggregation rule. Appendix C defines the property-level binary prediction as the OR of the three street-view predictions: 'If any of the three street views is predicted as positive, the property-level outcome is set to 1.' Under this rule, adding views mechanically increases recall and decreases precision/specificity, which is exactly the pattern reported in Fig. 3b. This does not by itself establish that multiple views contribute information. The authors should compare OR aggregation with alternatives (AND, majority vote, mean-probability threshold) and report single-view baselines under the same evaluation. Without this, claim (i) is not fully supported.
  4. [§3.4 and §4] All classification metrics are computed on a single 30% split with no confidence intervals, bootstrap resampling, or significance tests for accuracy, F1, and Kappa. Given that the ensemble's advantage over the best base model is modest and partially contradicted by the text, point estimates without uncertainty are insufficient. Please provide confidence intervals or paired tests (e.g., bootstrap over properties, McNemar's test) for at least accuracy, F1, and Kappa, and ideally for the threshold-tuning procedure as well.
minor comments (4)
  1. [Title] The title appears to contain a typo: 'Accessed' should likely be 'Assessed'.
  2. [§5.4] The phrase 'moderate agreement between prepositions and observations' should read 'between predictions and observations'.
  3. [§6, Conclusion] The statement 'it proved accurate compared to trained in-person assessors' overstates the results, given the paper's own report that Kappa is below 0.6 for the residential conditions.
  4. [References] Some citations are inconsistent, e.g., 'Zuo & Wang et al., 2022' appears where the second author's name is likely 'Wang'; please check the reference list for correctness.

Circularity Check

1 steps flagged

Ensemble advantage rests in part on a decision threshold fitted to the test labels, making the headline comparison partly an in-sample fit rather than an independent prediction.

specific steps
  1. fitted input called prediction [Section 3.3 (threshold selection) and Fig. 6 caption / Section 4.3 (test-set evaluation)]
    "Following training, predicted probabilities were converted into binary classifications by sweeping thresholds from 0.10 to 0.90 (in increments of 0.01). The threshold maximizing Cohen’s Kappa (McHugh, 2012) was selected ... The resulting model and tuned threshold were then fixed for all downstream ensemble predictions. ... Comparison between ensemble and base models based on the test set: the performance of the ensemble learner and base models with threshold selected for maximizing kappa (a) ..."

    The paper describes a 70/30 train/test split but no validation split for threshold selection, and the reported metrics are explicitly 'based on the test set.' The decision threshold is therefore optimized on the same labels used to compute the headline Kappa, F1, and accuracy values. The ensemble's reported advantage over the base models is thus partly a function of fitting a threshold to the test set, rather than an out-of-sample prediction. Since no analogous threshold search is reported for the base models, the comparison is lopsided and the central claim 'the ensemble learner outperforms individual base models' is partially forced by the fitting procedure.

full rationale

The core evaluation is not wholly circular: the VLM outputs are zero-shot, the ground truth is an independent professional survey, and the XGBoost meta-learner is trained on a separate 70% split. However, one load-bearing step is a fitted input presented as a prediction: §3.3 sweeps thresholds and selects the value maximizing Cohen's Kappa, and §4.3 reports test-set metrics from that same threshold selection, with the figure caption explicitly stating the comparison is 'based on the test set.' This makes the ensemble's Kappa/F1/accuracy values in-sample optima of the decision threshold, and the comparison to base models is not controlled for threshold tuning. The paper also notes in §4.3 a 'slight drop in accuracy and F1-score compared to Gemma3,' which conflicts with the abstract's unqualified 'outperforms individual base models' claim; that is a consistency/correctness issue rather than a circularity issue. No self-citation chain, imported uniqueness theorem, or ansatz-by-citation was found; the weighted blight rubric is stated as coming from city assessor feedback, not from the model outputs. Overall, the central ensemble-advantage claim is partly circular because one of its reported performance metrics is optimized on the test data, but the underlying framework remains independently benchmarked against external annotations.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central result depends on the supplied professional labels, a hand-chosen blight weighting, and a test-set threshold fit; no new physical or conceptual entities are introduced.

free parameters (3)
  • Blight score weights = roof +3, opening +1.5, facade +4
    Chosen from city assessor feedback (Appendix D), not derived from data; changes to these weights change the blight score comparisons.
  • Decision threshold (Kappa) = Selected per condition from 0.10-0.90 sweep to maximize test-set Kappa
    Tuned on the test set, which inflates reported ensemble performance (Section 3.3).
  • Decision threshold (F1) = Selected per condition from 0.10-0.90 sweep to maximize test-set F1
    Same issue as Kappa threshold; used for the high-recall variant.
axioms (4)
  • domain assumption Professional DLBA annotations are accurate ground truth for roof, opening, and facade damage
    The paper uses these as the reference labels for all metrics; if the original ordinal-to-binary collapse is noisy, reported accuracy is miscalibrated.
  • domain assumption Three street-view images per property provide sufficient visual coverage of the assessed conditions
    The models are scored only on these views; damage hidden from all three views cannot be detected, confounding model error with missing information.
  • domain assumption The weighting of blight components (roof +3, opening +1.5, facade +4) reflects the construct of blight severity
    The final blight score used for RMSE/MAE/CCC comparisons is an arbitrary weighted sum, not a validated measurement scale.
  • ad hoc to paper The 70/30 stratified split gives an unbiased evaluation when the threshold is selected on the test set
    The paper selects thresholds on the test set, which violates the standard assumption of a clean train/test separation; this is an ad hoc choice that biases the results.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Urban Blight Be Accessed with Vision-language Models: A Case Study in Detroit." pith.science (2026). https://pith.science/paper/RT5YUIRH

@misc{pith2026260801753,
  author       = {Pith},
  title        = {Pith review of: Can Urban Blight Be Accessed with Vision-language Models: A Case Study in Detroit},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RT5YUIRH}},
  note         = {Machine review of arXiv:2608.01753}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Addressing urban blight has seen increased focus in the past 15 years. Assessing urban blight is essential for guiding urban planning, targeting rehabilitation, and safeguarding public health, yet traditional residential blight surveys are difficult to maintain at scale due to the labor-intensive cost and long-term cycle. This study introduced a scalable framework for estimating residential blight using open-source large vision-language models on multiple views. Structured prompts guided models to evaluate housing attributes, including roof integrity, wall damage, and broken or boarded openings, producing both binary assessments and probabilistic estimates of disrepair. To evaluate the performance of these visual assessments, we compared professional human annotations of these features across several models, including an ensemble stacking approach based on XGBoost and a weighted scoring system. Results showed that (i) multiple street views can contribute to the improvement of accuracy, (ii) large vision-language models have different strengths of inference, (iii) the ensemble learner outperforms individual base models, enhancing robustness across all residential conditions and blight assessment. The practical application of the method allows low-cost tracking and management of housing stock conditions, providing a regularly updatable complement to traditional blight surveys.

Figures

Figures reproduced from arXiv: 2608.01753 by Aohua Tian, Derek Van Berkel, Mark Lindquist, Xiaohao Yang, Xu Qiang.

Figure 1
Figure 1. Figure 1: Framework overview of the method [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Case study areas: the tract-level site was selected based on demographic and land value data (left); the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Model performance with and without multiple views. (a) Model performance with multiple street views, and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Agreement on residential conditions (top) and average damage probability (bottom) between models (Note: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of (a) estimated blight and (b) estimation error [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between ensemble and base models based on the test set: the performance of the ensemble [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Error distribution (top) and concordance correlation coefficient (bottom): In concordance correlation [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: SHAP value of ensemble models (a: roof; b: opening; c: facade) [PITH_FULL_IMAGE:figures/full_fig_p011_8.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

63 extracted references · 53 canonical work pages · 1 internal anchor

  1. [1]

    (2018, June)

    Alam, F., Ofli, F., & Imran, M. (2018, June). Crisismmd: Multimodal twitter datasets from natural disasters. In Proceedings of the international AAAI conference on web and social media (V ol. 12, No. 1)

    work page 2018

  2. [2]

    Atkinson, R. (2004). The evidence on the impact of gentrification: new lessons for the urban renaissance?. European journal of housing policy, 4(1), 107-131

    work page 2004

  3. [3]

    & Lin, J

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., ... & Lin, J. (2025). Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923. 16 Can Urban Blight Be Accessed with Vision-language Models?A PREPRINT

    Pith/arXiv arXiv 2025

  4. [4]

    Beers, A., Daley, C., McLaughlin, I., & Pavlek, G. (2011). Quick guide: New tools to address blight and abandonment. Pennsylvania: The Housing Alliance of Pennsylvania

    work page 2011

  5. [5]

    Bibri, S. E. (2019). On the sustainability of smart and smarter cities in the era of big data: an interdisciplinary and transdisciplinary literature review. Journal of Big Data, 6(1), 25

    work page 2019

  6. [6]

    S., Chen, Q., & Engelmann, F

    Bieri, V ., Zamboni, M., Blumer, N. S., Chen, Q., & Engelmann, F. (2025, February). Opencity3d: What do vision-language models know about urban environments?. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV) (pp. 5147-5155). IEEE

    work page 2025

  7. [7]

    Bonnefoy, X. (2007). Inadequate housing and health: an overview. International journal of environment and pollution, 30(3-4), 411-429

    work page 2007

  8. [8]

    Board on Health Promotion, Disease Prevention, & Committee on Damp Indoor Spaces. (2004). Damp indoor spaces and health

    work page 2004

  9. [9]

    Buckland, M., & Gey, F. (1994). The relationship between recall and precision. Journal of the American society for information science, 45(1), 12-19

    work page 1994

  10. [10]

    Cao, J., Wang, H., Li, J., Tian, Q., & Niyogi, D. (2022). Improving the forecasting of winter wheat yields in Northern China with machine learning–dynamical hybrid subseasonal-to-seasonal ensemble prediction. Remote Sensing, 14(7), 1707

    work page 2022

  11. [11]

    Chai, T., & Draxler, R. R. (2014). Root mean square error (RMSE) or mean absolute error (MAE). Geoscientific model development discussions, 7(1), 1525-1534

    work page 2014

  12. [12]

    Chen, T., He, T., Benesty, M., & Khotilovich, V . (2019). Package ’xgboost’. R version, 90(1-66), 40

    work page 2019

  13. [13]

    & Zhou, T

    Chen, T., He, T., Benesty, M., Khotilovich, V ., Tang, Y ., Cho, H., ... & Zhou, T. (2015). Xgboost: extreme gradient boosting. R package version 0.4-2, 1(4), 1-4

    work page 2015

  14. [14]

    Chen, Z., Li, J., Chen, P., Li, Z., Sun, K., Luo, Y ., ... & Yu, P. S. (2025). Harnessing multiple large language models: A survey on llm ensemble. arXiv preprint arXiv:2502.18036

    Pith/arXiv arXiv 2025

  15. [15]

    Cho, D., Yoo, C., Im, J., Lee, Y ., & Lee, J. (2020). Improvement of spatial interpolation accuracy of daily maximum air temperature in urban areas using a stacking ensemble technique. GIScience & Remote Sensing, 57(5), 633-649

    work page 2020

  16. [16]

    Detroit Land Bank Authority. (2025). City Council quarterly report: Q1 FY25. Detroit Land Bank Authority. https://dlba-production-bucket.s3.us-east- 2.amazonaws.com/City_Council_Quarterly_Report/DLBA+Q1+FY25+CCQR+FINAL.pdf

    work page 2025

  17. [17]

    Dona, M. A. M., Cabrero-Daniel, B., Yu, Y ., & Berger, C. (2024). Evaluating and enhancing trustworthiness of LLMs in perception tasks. arXiv preprint arXiv:2408.01433. https://arxiv.org/abs/2408.01433

    Pith/arXiv arXiv 2024

  18. [18]

    Erb-Downward, J., & Merchant, S. (2020). Losing home: Housing instability & availability in Detroit. Retrieved from Poverty Solutions website: https://poverty. umich. edu/files/2020/05/200358_Poverty- Solutions_Detroit-Housing-Instability-policy-brief_051120. Pdf

    work page 2020

  19. [19]

    Feng, J., Du, Y ., Liu, T., Guo, S., Lin, Y ., & Li, Y . (2024). CityGPT: Empowering urban spatial cognition of large language models. arXiv preprint arXiv:2406.13948. https://arxiv.org/abs/2406.13948

    Pith/arXiv arXiv 2024

  20. [20]

    M., Behera, M

    Ghosh, S. M., Behera, M. D., Jagadish, B., Das, A. K., & Mishra, D. R. (2021). A novel approach for estimation of aboveground biomass of a carbon-rich mangrove site in India. Journal of Environmental Management, 292, 112816

    work page 2021

  21. [21]

    E., Ruffley, M., & Exposito-Alonso, M

    Gillespie, L. E., Ruffley, M., & Exposito-Alonso, M. (2024). Deep learning models map rapid plant species changes from citizen science and remote sensing data. Proceedings of the National Academy of Sciences, 121(37), e2318296121

    work page 2024

  22. [22]

    P., Cohen, W

    Healey, S. P., Cohen, W. B., Yang, Z., Brewer, C. K., Brooks, E. B., Gorelick, N., ... & Zhu, Z. (2018). Mapping forest change using stacked generalization: An ensemble approach. Remote Sensing of Environment, 204, 717-728

    work page 2018

  23. [23]

    Huang, Z., Qi, H., Kang, C., Su, Y ., & Liu, Y . (2020). An ensemble learning approach for urban land use mapping based on remote sensing imagery and social sensing data. Remote Sensing, 12(19), 3254

    work page 2020

  24. [24]

    Jiang, Y ., Chao, Q., Chen, Y ., Li, X., Liu, S., & Cong, G. (2024). UrbanLLM: Autonomous ur- ban activity planning and management with large language models. arXiv preprint arXiv:2406.12360. https://arxiv.org/abs/2406.12360

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    Jiang, Y ., & Sun, P. (2024). Does shrinkage have an impact on urban livability? An empirical analysis from Northeast China. Sustainable Cities and Society, 113, 105725. 17 Can Urban Blight Be Accessed with Vision-language Models?A PREPRINT

    work page 2024

  26. [26]

    S., Reid, M., Matsuo, Y ., & Iwasawa, Y

    Kojima, T., Gu, S. S., Reid, M., Matsuo, Y ., & Iwasawa, Y . (2022). Large language models are zero-shot reasoners. Advances in neural information processing systems, 35, 22199-22213

    work page 2022

  27. [27]

    Lawrence, I., & Lin, K. (1989). A concordance correlation coefficient to evaluate reproducibility. Biometrics, 255-268

    work page 1989

  28. [28]

    Li, Y ., & Long, Y . (2024). Inferring storefront vacancy using mobile sensing images and computer vision approaches. Computers, Environment and Urban Systems, 108, 102071

    work page 2024

  29. [29]

    Li, Z., Xia, L., Tang, J., Xu, Y ., Shi, L., Xia, L., Yin, D., & Huang, C. (2024). UrbanGPT: Spatio-temporal large language models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 5351–5362). Association for Computing Machinery. https://doi.org/10.1145/3637528.3671578

    work page arXiv 2024

  30. [30]

    Liang, X., Brainerd, B., Hicks, T., & Andris, C. (2024). Lessons from a human-in-the-loop machine learning approach for identifying vacant, abandoned, and deteriorated properties in Savannah, Georgia. Journal of Planning Education and Research, 0739456X241273945

    work page 2024

  31. [31]

    Liang, X., Xie, J., Zhao, T., Stouffs, R., & Biljecki, F. (2025). OpenFACADES: An Open Framework for Ar- chitectural Caption and Attribute Data Enrichment via Street View Imagery. arXiv preprint arXiv:2504.02866

    arXiv 2025

  32. [32]

    Lv, L., Chen, T., Dou, J., & Plaza, A. (2022). A hybrid ensemble-based deep-learning framework for landslide susceptibility mapping. International Journal of Applied Earth Observation and Geoinformation, 108, 102713

    work page 2022

  33. [33]

    Malekzadeh, M., Willberg, E., Torkko, J., & Toivonen, T. (2025). Urban attractiveness according to ChatGPT: Contrasting AI and human insights. Computers, Environment and Urban Systems, 117, 102243

    work page 2025

  34. [34]

    McHugh, M. L. (2012). Interrater reliability: the kappa statistic. Biochemia medica, 22(3), 276-282

    work page 2012

  35. [35]

    (2025, March 17)

    Mistral AI. (2025, March 17). Mistral Small 3.1. Mistral AI. https://mistral.ai/news/mistral-small-3-1

    work page 2025

  36. [36]

    Neidert, L., Farley, R., & Morenoff, J. (2025). How Census Undercount Became a Civil Rights Issue and Why It Is Increasingly Important. RSF: The Russell Sage Foundation Journal of the Social Sciences, 11(1), 26–43

    work page 2025

  37. [37]

    (2025, July)

    Niimi, J. (2025, July). A simple ensemble strategy for llm inference: Towards more stable text classification. In International Conference on Applications of Natural Language to Information Systems (pp. 189-199). Cham: Springer Nature Switzerland

    work page 2025

  38. [38]

    M., Moturi, W

    Ochodo, C., Ndetei, D. M., Moturi, W. N., & Otieno, J. O. (2014). External built residential environment characteristics that affect mental health of adults. Journal of Urban Health, 91, 908-927

    work page 2014

  39. [39]

    Peng, Q., Zhao, G., & Ye, X. (2025). Assessing the impact of maintenance condition on multifamily rents: an integrated approach of machine learning and hedonic modelling. Journal of Housing and the Built Environment, 1-18

    work page 2025

  40. [40]

    J., Reeves, A., Baker, E., & Bentley, R

    Pevalin, D. J., Reeves, A., Baker, E., & Bentley, R. (2017). The impact of persistent poor housing conditions on mental health: A longitudinal population-based study. Preventive medicine, 105, 304-310

    work page 2017

  41. [41]

    M., Ferreira, F

    Pinto, A. M., Ferreira, F. A., Spahr, R. W., Sunderman, M. A., Govindan, K., & Meidut˙e-Kavaliauskien˙e, I. (2021). Analyzing blight impacts on urban areas: A multi-criteria approach. Land Use Policy, 108, 105661

    work page 2021

  42. [42]

    M., Ferreira, F

    Pinto, B. M., Ferreira, F. A., Spahr, R. W., Sunderman, M. A., & Pereira, L. F. (2023). Analyzing causes of urban blight using cognitive mapping and DEMATEL. Annals of operations research, 325(2), 1083-1110

    work page 2023

  43. [43]

    Ruggiero, R., Rivera, J., & Cooney, P. (2020). A decent home: The status of home repair in Detroit. Ann Arbor: Poverty Solutions University of Michigan

    work page 2020

  44. [44]

    Sagi, O., & Rokach, L. (2018). Ensemble learning: A survey. Wiley interdisciplinary reviews: data mining and knowledge discovery, 8(4), e1249

    work page 2018

  45. [45]

    & Iqbal, S

    Team, G., Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., ... & Iqbal, S. (2025). Gemma 3 technical report. arXiv preprint arXiv:2503.19786

    Pith/arXiv arXiv 2025

  46. [46]

    M., & Witten, I

    Ting, K. M., & Witten, I. H. (1999). Issues in stacked generalization. Journal of artificial intelligence research, 10, 271-289

    work page 1999

  47. [47]

    Trevethan, R. (2017). Sensitivity, specificity, and predictive values: foundations, pliabilities, and pitfalls in research and practice. Frontiers in public health, 5, 307

    work page 2017

  48. [48]

    American Community Survey 2018 1-year estimates [Dataset and codebook]

    US Bureau of the Census (2018). American Community Survey 2018 1-year estimates [Dataset and codebook]. Retrieved from https://censusreporter.org/profiles/16000US2622000-detroit-mi/

    work page 2018

  49. [49]

    Congress

    U.S. Congress. (2021). Exploring How Community Development Financial Institutions Support Underserved Communities: Hearing Before the Committee on Banking, Housing, and Urban Affairs, United States Senate, One Hundred Seventeenth Congress, First Session. U.S. Senate Committee on Banking, Housing, and Urban Affairs. 18 Can Urban Blight Be Accessed with Vis...

    work page 2021

  50. [50]

    Verma, D., Mumm, O., & Carlow, V . M. (2023). Generative agents in the streets: Exploring the use of large language models (LLMs) in collecting urban perceptions. arXiv preprint arXiv:2312.13126. https://arxiv.org/abs/2312.13126

    arXiv 2023

  51. [51]

    tidycensus: Load US Census Boundary and Attribute Data as ’tidyverse’ and ’sf’-Ready Data Frames

    Walker K, Herman M (2025). tidycensus: Load US Census Boundary and Attribute Data as ’tidyverse’ and ’sf’-Ready Data Frames. R package version 1.7.1, https://walker-data.com/tidycensus/

    work page 2025

  52. [52]

    Wang, Z., Majumdar, A., & Rajagopal, R. (2023). Geospatial mapping of distribution grid with machine learning and publicly-accessible multi-modal data. Nature Communications, 14(1), 5006

    work page 2023

  53. [53]

    & Zhou, D

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., ... & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35, 24824-24837

    2022

  54. [54]

    Wolpert, D. H. (1992). Stacked generalization. Neural networks, 5(2), 241-259

    work page 1992

  55. [55]

    H., & Macready, W

    Wolpert, D. H., & Macready, W. G. (2002). No free lunch theorems for optimization. IEEE transactions on evolutionary computation, 1(1), 67-82

    work page 2002

  56. [56]

    Yan, Y ., Wen, H., Zhong, S., Chen, W., Chen, H., Wen, Q., Zimmermann, R., & Liang, Y . (2024). UrbanCLIP: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web. In Proceedings of the ACM Web Conference 2024 (pp. 4006–4017). Association for Computing Machinery. https://doi.org/10.1145/3589334.3645378

    arXiv 2024

  57. [57]

    Yan, Y ., Zeng, Q., Zheng, Z., Yuan, J., Feng, J., Zhang, J., Xu, F., & Li, Y . (2024). OpenCity: A scal- able platform to simulate urban activities with massive LLM agents. arXiv preprint arXiv:2410.21286. https://arxiv.org/abs/2410.21286

    Pith/arXiv arXiv 2024

  58. [58]

    (1999, August)

    Yang, Y ., & Liu, X. (1999, August). A re-examination of text categorization methods. In Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval (pp. 42-49)

    work page 1999

  59. [59]

    Yue, X., Wang, Y ., Zhao, Y ., & Zhang, H. (2022). Estimation of urban housing vacancy based on daytime housing exterior images–a case study of Guangzhou in China. ISPRS International Journal of Geo-Information, 11(6), 349

    work page 2022

  60. [60]

    S., & Yin, K

    Zeng, T., Wu, L., Peduto, D., Glade, T., Hayakawa, Y . S., & Yin, K. (2023). Ensemble learning framework for landslide susceptibility mapping: Different basic classifier and ensemble strategy. Geoscience Frontiers, 14(6), 101645

    work page 2023

  61. [61]

    Zhang, Y ., Liu, J., & Shen, W. (2022). A review of ensemble learning algorithms used in remote sensing applications. Applied Sciences, 12(17), 8654

    work page 2022

  62. [62]

    Zhou, Z. H. (2025). Ensemble methods: foundations and algorithms. CRC press

    work page 2025

  63. [63]

    Zou, S., & Wang, L. (2022). Mapping individual abandoned houses across cities by integrating VHR remote sensing and street view imagery. International Journal of Applied Earth Observation and Geoinformation, 113, 103018. 19

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.