Life Style Levels: Neighborhood Delineation using Geospatial Data

Debarag Banerjee; Srivatsa Kulkarni

arxiv: 2607.06529 · v1 · pith:QFOJIKQY · submitted 2026-07-07 · cs.CL

Life Style Levels: Neighborhood Delineation using Geospatial Data

Srivatsa Kulkarni , Debarag Banerjee This is my paper

Reviewed by Pith2026-07-08 02:43 UTCglm-5.2pith:QFOJIKQY open to challenge →

classification cs.CL

keywords urban delineationbuilding morphologyaffluence mappingDBSCAN clusteringOpen Buildings datasetinformal settlementssatellite imagerycredit delinquency

0 comments

The pith

Building Footprints Predict Indian Neighborhood Affluence

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

The paper claims that a single rule-based score — median building area divided by building count, computed from open-source satellite footprints at 100-meter resolution — can classify Indian urban neighborhoods into affluence tiers that correspond to real socioeconomic differences. The authors partition 59 cities into uniform grids, score each grid with this ratio, and label the lowest-scoring 5% as 'Red' (dense, likely disadvantaged), the next 4% as 'Amber' (transitional), and the rest as 'Green' (planned, spacious). They validate these labels through Google Street View imagery and by showing that DBSCAN clustering of building centroids in Mumbai overlaps with known informal settlements. As a final check, they map consumer loan delinquency onto the grid classes and find a monotonic ordering — Red above Amber above Green — in 26 of 59 cities.

Core claim

The central mechanism is the Grid Score: Grid Score = Median(Building Area) / Building Count. This ratio exploits an observed inverse relationship between how many buildings pack into a cell and how large those buildings are. Dense informal settlements produce many small footprints, yielding low scores; planned neighborhoods produce fewer, larger footprints, yielding high scores. The paper's contribution is showing that this single transparent ratio, applied uniformly across diverse Indian cities using only free satellite data, produces neighborhood classes that align with visual ground truth and correlate with financial behavior.

What carries the argument

Grid Score = Median(Building Area) / Building Count, computed per 100m × 100m cell from the Open Buildings dataset; percentile-based thresholds (5th, 9th) for Red/Amber/Green classification; DBSCAN clustering of building centroids with Eps=11m, MinPts=5 for informal settlement detection in Mumbai

If this is right

Any organization with access to free satellite building footprints could produce neighborhood-level affluence maps for Indian cities without census data, household surveys, or commercial imagery.
The delinquency correlation, where it holds, suggests that building morphology alone carries signal relevant to credit risk assessment, potentially supplementing traditional underwriting in data-sparse markets.
The DBSCAN-based informal settlement detection in Mumbai, if generalized, could provide an automated, label-free method for identifying deprived areas across cities where slum maps do not exist.
The framework's failure in rural areas — where uniform housing density collapses the score distribution — highlights a boundary condition: the method requires morphological heterogeneity to function.

Where Pith is reading between the lines

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

The monotonic delinquency trend holding in only 26 of 59 cities suggests the building-count-to-size ratio may track affluence more reliably in larger, more morphologically heterogeneous cities than in smaller towns where building patterns are more uniform.
If 2.5D building height data were incorporated, the score could distinguish between dense low-rise slums and dense high-rise middle-class apartments — two cases that currently produce similar low Grid Scores but represent very different affluence levels.
The percentile-based thresholding (5th and 9th) implicitly assumes every city has the same proportion of disadvantaged neighborhoods; cities with atypically high or low deprivation shares would be misclassified at the margins.
The method could be tested against Indian census ward-level data where available, or against nighttime light intensity, to establish whether the Grid Score adds signal beyond what simpler proxies already capture.

Load-bearing premise

The load-bearing premise is that fewer, larger buildings in a 100-meter cell reliably indicate greater affluence while many smaller buildings indicate deprivation — a relationship the paper validates only through Street View images and a delinquency correlation that holds in fewer than half the cities studied, without direct income or wealth data to confirm the link.

What would settle it

Find a city where dense middle-class apartment complexes (many buildings, small footprints) score as 'Red' or where large warehouse or industrial structures (few buildings, large footprints) score as 'Green,' breaking the affluence mapping.

read the original abstract

Fine-scale socioeconomic information is often unavailable across rapidly ur-banizing regions of the developing world, like India, limiting the ability to delineate intra-urban variations in affluence and deprivation. This study pro-poses a scalable, grid-based urban delineation framework using building morphology derived from open-source satellite imagery. Urban areas across 59 Indian cities and towns are partitioned into high-resolution spatial grids and characterized using interpretable morphological indicators, which are combined into a transparent, rule-based scoring framework to delineate areas with contrasting levels of urban affluence. The resulting classifications are validated through ground-level Google Street View observations, revealing a sharp contrast between the grid classes which are consistent with the ex-pected effects of the lifestyle affluence indicators. We further investigate density-based clustering of building footprints in Mumbai to identify dense urban settlements, demonstrating that the resulting clusters exhibit substan-tial spatial overlap with known informal settlements across the city. Finally, we conduct an exploratory analysis mapping consumer loan delinquency across the derived affluence classes. By relying entirely on publicly available geospatial data, the proposed framework provides a scalable, interpretable, and cost-effective approach for granular urban affluence mapping across In-dian cities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Paper applies a simple, parameter-free grid score to 59 Indian cities as an affluence proxy, but the formula has a structural confound and validation is almost entirely qualitative.

read the letter

Bottom line: this paper proposes a parameter-free grid score — Median(Building Area) / Building Count — computed from open building footprints at 100m resolution across 59 Indian cities, then classifies grids into Red/Amber/Green affluence tiers and checks whether delinquency rates track those tiers. The formula itself is clean and reproducible from public data, and the scale of application (59 cities) is the genuinely new contribution. The DBSCAN clustering on Mumbai building centroids, with visual confirmation of overlap with known informal settlements like Dharavi, is a reasonable exploratory exercise and honestly presented as such. The authors deserve credit for using entirely open data and for not overselling their predictive model — they report a Gini of 14–15% and call it weak, which is accurate. The parameter-free nature of the core score is a real strength; no learned weights, no fitting. The Street View images in the appendix do show visible morphological contrast between Red and Green grids. That said, the soft spots are substantial and mostly load-bearing. The stress-test concern about the formula conflating dense affluent apartment complexes with informal settlements is valid and not fully addressed. The Purple class (≤2 buildings) catches sparse non-residential grids, but a grid with 3+ warehouses or a dense middle-class apartment block would still score low and get labeled Red. The percentile-based classification forces every city to have exactly 5% Red grids regardless of actual deprivation levels — this is a design choice, not a discovery, and the paper's language about 'statistical consistency' across cities conflates an artifact of the method with evidence about the world. The delinquency validation is the only quantitative check, and it's weak: the monotonic trend holds in 26 of 59 cities, and several cities (Cuttack 0.758, Berhampore 0.399, Bhavnagar 0.782) show Red grids with delinquency below the city average. No ground-truth income or wealth data is used anywhere. DBSCAN parameters (eps=11, MinPts=5) were tuned on Mumbai alone and applied nowhere else. No code or boundary data is released. The reader's CONDITIONAL verdict and the stress-test concerns both land. I'd add that the reader slightly over-scored circularity (4.0) — the grid score formula itself is genuinely parameter-free, so the circularity is really in the threshold calibration, not the core metric. This is an exploratory framework with a real idea at its center, but it needs quantitative validation against any external socioeconomic ground truth before the central claim can be taken seriously. Recommend serious peer review — the core idea is worth engaging with, and a good referee would push the authors toward the validation and sensitivity analysis that would either confirm or constrain the approach.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes a grid-based framework for delineating urban neighborhood affluence across 59 Indian cities using the open-source Open Buildings dataset. At 100m × 100m resolution, each grid is assigned a rule-based Grid Score defined as Median(Building Area) / Building Count. Grids are classified into Red (deprived), Amber (transitional), Green (affluent), and Purple (non-residential) based on percentile thresholds. The authors validate the classes qualitatively using Google Street View, demonstrate that DBSCAN clustering of building footprints in Mumbai overlaps with known informal settlements, and show an exploratory monotonic relationship between grid class and consumer loan delinquency in 26 of 59 cities.

Significance. The pursuit of scalable, interpretable, and cost-free proxies for intra-urban socioeconomic variation is well-motivated, particularly for developing regions where fine-scale census data is scarce. The parameter-free formulation of the Grid Score (Eq. 1) is a strength in its simplicity and reproducibility. The integration of building morphology with financial delinquency data provides a falsifiable downstream validation signal. However, the significance of these findings is currently undermined by structural confounds in the core metric and the absence of quantitative validation.

major comments (4)

§4.2.2, Eq. (1): The Grid Score formula (Median Building Area / Building Count) has a structural confound that is load-bearing for the central claim. By placing Building Count in the denominator without any land-use filter, the formula penalizes dense residential areas regardless of their affluence. A 100m grid containing multiple middle-class apartment buildings (e.g., median area 150 m², count=15) scores 10 and is labeled 'Red', while a grid with a single non-residential warehouse (area 2000 m², count=1) scores 2000 and is labeled 'Green'. The 'Purple' classification for grids with ≤2 buildings (§4.2.2) does not resolve this for grids with 3+ non-residential structures. The authors must address how this confound affects the validity of the affluence proxy, particularly for dense but affluent residential complexes common in Indian cities.
§4.2.2: The percentile-based classification forces exactly 5% of grids to be labeled 'Red' and 4% 'Amber' in every city, regardless of the actual prevalence of deprivation. The authors state these thresholds were derived from 'manual inspection of a large random sample of grids across nine representative cities.' This calibration on the same data being classified introduces a risk of circularity. A uniformly affluent city would still have 5% of its grids labeled 'deprived.' The authors need to justify why a fixed percentile is appropriate across cities with vastly different socioeconomic profiles, or demonstrate sensitivity of the delinquency results to alternative thresholding schemes.
§5.1, Table 7: The delinquency validation is cited as supporting the framework, but the monotonic trend (Red > Amber > Green) holds in only 26 of 59 cities (44%). In 33 cities, the expected ordering does not hold. This is below the threshold one would expect if the Grid Score reliably tracked affluence. The manuscript does not analyze why the trend fails in the majority of cities, nor does it discuss whether the structural confound (dense affluent areas misclassified as 'Red') could explain these failures. This proportionally weakens the claim that the framework generalizes across diverse Indian cities.
§5.2: The validation of the affluence classes is entirely qualitative, relying on Google Street View images without any quantitative accuracy metric, inter-rater reliability assessment, or systematic sampling protocol. While illustrative, these images cannot support a claim of general validity across 59 cities. The addition of a quantitative ground-truth comparison (e.g., against census ward-level income data where available, or a systematic human-annotated sample) is needed to substantiate the central claim.

minor comments (6)

§4.4, Table 2 vs. §4.4 text: The AREA_BIN thresholds in Table 2 are defined as <40 m² (LOW), 40–190 m² (MEDIUM), and >190 m² (HIGH), but the text in §4.4 describes different thresholds: <50 m², 50–120 m², and 120–200 m². These should be reconciled.
§4.3: The DBSCAN parameters (Eps=11, MinPts=5) and the cluster score threshold of 8000 (Table 5) were tuned specifically for Mumbai. The manuscript states the clustering study was 'mainly done for Mumbai city,' but the abstract and introduction imply a broader framework. The scope of the DBSCAN claims should be clarified.
§4.4, Eq. (3): The logistic regression assumes 'Independence of Observations: Credit behavior of a person in a grid is not influenced by a neighboring grid.' Given the spatial nature of the data, spatial autocorrelation is a plausible violation. This assumption should be discussed.
§5.1, Table 8: The definition of 'Close Trend' includes D(Green)=D(Amber)<D(Red), but this is actually a partial monotonic trend. The rationale for treating this as a separate, less stringent category should be clarified.
References: Several citations appear to have formatting issues or future dates (e.g., Refs 9, 10, 13, 19 accessed in 2026). Reference 20 is cited as 'Siddiqi N. Credit Risk Scorecards' but is listed as source [20] for both the concave hull and logistic regression. The numbering should be checked.
§6, Conclusions: The phrase 'A strong positive correlation (can be positive or negative) of delinquency' is internally contradictory and should be revised.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and substantive review. The referee raises four major concerns: (1) a structural confound in the Grid Score formula that penalizes dense affluent residential areas; (2) potential circularity from fixed-percentile thresholds applied uniformly across cities; (3) the delinquency trend holding in only 26 of 59 cities without analysis of failures; and (4) the absence of quantitative validation. We agree that all four points identify genuine weaknesses that require revision. We will address the structural confound by introducing a land-use filter and exploring alternative score formulations; we will add sensitivity analysis for alternative thresholding schemes; we will analyze the cities where the trend fails and discuss the confound as a possible explanation; and we will add a quantitative validation component using census ward-level data where available and a systematic human-annotated sample with inter-rater reliability. We cannot fully resolve the fixed-percentile circularity concern without absolute ground-truth deprivation data across all 59 cities, which is not available, but we will transparently acknowledge this limitation and provide sensitivity results.

read point-by-point responses

Referee: §4.2.2, Eq. (1): The Grid Score formula (Median Building Area / Building Count) has a structural confound that is load-bearing for the central claim. By placing Building Count in the denominator without any land-use filter, the formula penalizes dense residential areas regardless of their affluence. A 100m grid containing multiple middle-class apartment buildings (e.g., median area 150 m², count=15) scores 10 and is labeled 'Red', while a grid with a single non-residential warehouse (area 2000 m², count=1) scores 2000 and is labeled 'Green'. The 'Purple' classification for grids with ≤2 buildings does not resolve this for grids with 3+ non-residential structures. The authors must address how this confound affects the validity of the affluence proxy, particularly for dense but affluent residential complexes common in Indian cities.

Authors: The referee is correct that the Grid Score formula as stated has a structural confound: by placing Building Count in the denominator without any land-use filter, dense residential areas—including affluent apartment complexes—can receive low scores and be misclassified as 'Red.' The referee's example of a grid with 15 middle-class apartment buildings scoring 10 is a valid illustration of this problem. We acknowledge this is a genuine weakness in the current formulation. In the revised manuscript, we will address this in three ways. First, we will introduce a pre-filtering step that uses OpenStreetMap land-use tags (where available) to exclude or separately flag grids dominated by non-residential structures (warehouses, industrial buildings, etc.) before computing the Grid Score, so that large non-residential footprints do not inflate scores into the 'Green' band. Second, we will explore an alternative formulation that decouples density from footprint size—for example, using Median Building Area alone as the primary score, with Building Count as a secondary modifier rather than a divisor—and will report comparative results. Third, we will explicitly discuss the limitation that dense affluent residential complexes (e.g., multi-story apartment blocks in cities like Mumbai and Bengaluru) may still be misclassified under the current and proposed formulations, and will identify this as a direction where building height data (e.g., from Open Buildings 2.5D Temporal) would be needed to resolve the ambiguity between vertical density and horizontal deprivation. We agree this confound is load-bearing and will revise the manuscript accordingly. revision: yes
Referee: §4.2.2: The percentile-based classification forces exactly 5% of grids to be labeled 'Red' and 4% 'Amber' in every city, regardless of the actual prevalence of deprivation. The authors state these thresholds were derived from 'manual inspection of a large random sample of grids across nine representative cities.' This calibration on the same data being classified introduces a risk of circularity. A uniformly affluent city would still have 5% of its grids labeled 'deprived.' The authors need to justify why a fixed percentile is appropriate across cities with vastly different socioeconomic profiles, or demonstrate sensitivity of the delinquency results to alternative thresholding schemes.

Authors: The referee raises a valid concern about circularity. The fixed-percentile approach was adopted because we observed empirical consistency in the proportion of high-density grids across the nine calibration cities, and because absolute deprivation ground truth is not available at the grid level across Indian cities. However, we agree that applying a fixed percentile forces every city to have 5% 'Red' grids regardless of actual deprivation prevalence, which is a substantive limitation. In the revised manuscript, we will address this by: (1) adding a sensitivity analysis in which we vary the Red threshold (e.g., 3rd, 5th, 7th, and 10th percentiles) and recompute the delinquency trend results across all 59 cities to demonstrate robustness or lack thereof; (2) exploring an alternative absolute-threshold scheme based on raw Grid Score values derived from the calibration cities, applied uniformly across all cities, and comparing the resulting class distributions; and (3) explicitly acknowledging in the limitations section that without city-level ground-truth deprivation data, we cannot fully resolve the circularity concern, and that the fixed-percentile approach should be interpreted as a relative within-city ranking rather than an absolute measure of deprivation. We note that the Korba rural extension example in §5.3 already illustrates the failure mode the referee describes—when the percentile approach is applied to areas with uniformly low density, all grids are classified as 'Green,' masking socioeconomic variation. We will expand this discussion. revision: yes
Referee: §5.1, Table 7: The delinquency validation is cited as supporting the framework, but the monotonic trend (Red > Amber > Green) holds in only 26 of 59 cities (44%). In 33 cities, the expected ordering does not hold. This is below the threshold one would expect if the Grid Score reliably tracked affluence. The manuscript does not analyze why the trend fails in the majority of cities, nor does it discuss whether the structural confound (dense affluent areas misclassified as 'Red') could explain these failures. This proportionally weakens the claim that the framework generalizes across diverse Indian cities.

Authors: The referee is correct that 26 out of 59 cities (44%) is a modest success rate and that the manuscript currently does not analyze the failures. This is a fair criticism. We will address it in the revision by: (1) conducting a systematic analysis of the 33 cities where the monotonic trend does not hold, examining whether these cities share common characteristics (e.g., smaller city size, lower customer sample sizes, different urban morphology patterns, or prevalence of dense affluent apartment complexes that the structural confound would misclassify as 'Red'); (2) explicitly testing whether the structural confound identified in the referee's first comment explains a subset of the failures—specifically, cities with a high proportion of multi-story residential complexes (e.g., Delhi, Noida, Gurugram) where dense affluent housing may be misclassified; (3) reporting customer sample sizes per city, as some of the 59 cities have as few as 10,000 customers, which may produce unstable delinquency rate estimates at the grid-class level; and (4) revising the language in the manuscript to accurately characterize the delinquency analysis as exploratory and to avoid overstating the generalizability claim. We agree that the current framing overstates the strength of the evidence, and we will temper the claims accordingly. revision: yes
Referee: §5.2: The validation of the affluence classes is entirely qualitative, relying on Google Street View images without any quantitative accuracy metric, inter-rater reliability assessment, or systematic sampling protocol. While illustrative, these images cannot support a claim of general validity across 59 cities. The addition of a quantitative ground-truth comparison (e.g., against census ward-level income data where available, or a systematic human-annotated sample) is needed to substantiate the central claim.

Authors: The referee is correct that the current validation is purely qualitative and does not meet the standard for supporting a general validity claim across 59 cities. We agree that quantitative validation is needed. In the revised manuscript, we will add: (1) a systematic human-annotated validation sample, in which a stratified random sample of grids (across Red, Amber, Green, and Purple classes, drawn from multiple cities) will be independently labeled by two annotators using Google Street View and satellite imagery, with inter-rater agreement reported (e.g., Cohen's kappa); (2) a quantitative comparison against census ward-level socioeconomic indicators (e.g., literacy rate, workforce participation, assets index from the 2011 Census of India) where ward boundaries can be mapped to grid classifications, computing correlation or rank-based statistics; and (3) a clear sampling protocol describing how grids were selected, how many were annotated, and what labeling criteria were used. We acknowledge that the 2011 Census data is outdated for some rapidly changing cities, and we will discuss this as a limitation. We cannot provide a comprehensive ground-truth comparison for all 59 cities because fine-scale socioeconomic data is not publicly available at sub-ward resolution across India, but we will provide a quantitative validation for a representative subset and will adjust the manuscript's claims to reflect the strength of evidence accurately. revision: yes

Circularity Check

2 steps flagged

One secondary validation (DBSCAN overlap with informal settlements) is partly forced by parameter tuning; the central Grid Score is parameter-free and validated against external delinquency data.

specific steps

fitted input called prediction [§4.3, DBSCAN parameter selection and §5/§4.3 validation of cluster overlap with informal settlements]
"Following iterative experimentation, manual cluster inspection, and parameter fine-tuning, we established values of Eps = 11 and MinPts = 5 for Mumbai. [...] many clusters in Mumbai demonstrate a striking overlap with informal settlements, Figure 12 to 16. This was verified via satellite imagery by the presence of thatched or tin roofs and closely packed houses."

The DBSCAN parameters (Eps=11, MinPts=5) were established through 'iterative experimentation, manual cluster inspection, and parameter fine-tuning' on Mumbai data. The paper then presents the overlap of the resulting clusters with known informal settlements as a finding ('demonstrate a striking overlap'). If the parameters were tuned by inspecting whether clusters matched known informal settlement locations, the overlap is partly a consequence of the tuning, not an independent discovery. The paper does not state that parameter selection was blind to informal settlement locations. However, this is a secondary analysis ('We further investigate...'), not the central Grid Score claim, so the circularity is limited in scope.
fitted input called prediction [§4.2.2, threshold derivation and percentile-based classification]
"To standardize these thresholds, we conducted a manual inspection of a large random sample of grids across nine representative cities to define the thresholds for the Red, Amber, and Green classes. Interestingly, our analysis revealed that Red grids consistently accounted for no more than 5% of the total urban area, while Amber grids never exceeded 9%. [...] we applied a percentile-based classification for the remaining 50 cities: grids within the 5th percentile are labelled Red, those up to the 9th percentile are labeled Amber, and all subsequent grids are labelled Green."

The 5th and 9th percentile cutoffs were derived from manual inspection of nine cities' grids (where the inspectors presumably identified which grids looked 'Red' vs 'Amber' vs 'Green' via Street View or satellite imagery). The paper then applies these same percentile cutoffs to 50 other cities and reports that the classification 'validates' well via Street View. The Street View validation on the 50 cities uses the same qualitative method (visual inspection) that was used to derive the thresholds on the 9 calibration cities. This is a standard fit-on-subset, apply-to-rest methodology, not strictly circular, but the 'interesting' finding that Red = 5% is an artifact of choosing the 5th percentile as the cutoff. The delinquency validation (§5.1) is genuinely external and mitigates this.

full rationale

The central Grid Score formula (Eq. 1: Median(Building Area) / Building Count) is parameter-free and not defined in terms of its own outputs. The main external validation—consumer loan delinquency from L&T Finance—is independent data not used in constructing the score. Two secondary analyses show mild circularity: (1) DBSCAN parameters for Mumbai were tuned by 'manual cluster inspection' and then the overlap with known informal settlements is presented as a finding, and (2) the 5th/9th percentile classification thresholds were derived from manual visual inspection of nine cities and then 'validated' using the same visual inspection method on other cities. Neither of these undermines the central claim, which rests on the parameter-free formula and the external delinquency correlation (observed in 26 of 59 cities). The circularity is real but limited to secondary validation steps.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, particles, or theoretical constructs. The free parameters are all data-processing thresholds tuned by manual inspection rather than principled selection. The domain assumptions are the most consequential: the density-footprint affluence proxy is the foundation of the entire framework, and its validity is established only through qualitative visual validation, not quantitative ground-truth comparison.

free parameters (6)

Percentile thresholds (5th, 9th) = 5th and 9th percentile of grid scores
Calibrated by manual inspection of nine cities; no statistical or theoretical justification provided for these specific percentiles.
DBSCAN eps = 11 meters
Tuned by iterative experimentation and manual cluster inspection on Mumbai only; no principled selection method.
DBSCAN MinPts = 5
Tuned alongside eps by manual inspection; no sensitivity analysis provided.
Cluster score threshold = 8000
Set after analyzing cluster scores in Mumbai to filter planned layouts; city-specific and ad hoc.
AREA_BIN thresholds = 40, 190 m² (and 50, 120, 200 m² for delinquency model)
Set by 'intuitive estimation of average house size in a metropolitan city'; two different threshold sets are used in different parts of the paper without reconciliation.
Address match threshold = 0.7 Levenshtein, score > 50
Chosen for geocoding quality filtering; no validation of impact on downstream results.

axioms (5)

domain assumption Building density inversely correlates with building footprint size as a proxy for neighborhood affluence
Invoked in §4.2.2: 'planned areas are characterized by a lower relative building count and have footprints of larger size and informal settlements have higher building count and smaller building footprint.' This is the foundational assumption of the grid score formula.
ad hoc to paper Percentile thresholds calibrated on nine Indian cities generalize to all 59 cities
The 5th and 9th percentile cutoffs derived from manual inspection of nine cities are applied uniformly to all 59 cities without testing cross-city generalization.
domain assumption Google Street View imagery is a valid ground truth for affluence classification
Used throughout §5.1-5.2 for validation. No quantitative accuracy metric is computed; validation is purely qualitative visual inspection.
domain assumption Loan delinquency rate is a valid proxy for neighborhood affluence
Used in §5.1 to validate the affluence classes. Delinquency is influenced by many factors beyond affluence (loan terms, borrower selection, economic shocks), and the correlation holds in only 26/59 cities.
ad hoc to paper Independence of observations across grids for logistic regression
Stated in §4.4 as a model assumption, but spatial autocorrelation between adjacent grids likely violates this assumption. No test for spatial independence is provided.

pith-pipeline@v1.1.0-glm · 17766 in / 2956 out tokens · 224638 ms · 2026-07-08T02:43:53.484716+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · 2 internal anchors

[1]

developed a cost -effective approach to derive morphological features from open - source Sentinel-2 imagery, using XGBoost [7] models to identify informal settlements in Africa. Their findings reveal that these settlements possess heterogeneous local char- acteristics, supporting the hypothesis of varying lifestyle levels, and they developed a generalized...

work page
[2]

AREA_BIN_

analyzed factors affecting nighttime lighting patterns, using brightness -based clus- tering to identify commercial and industrial zones. Their work established significant differences in brightness across various functional zones, including residential, under- developed, and industrial areas. Sungwon Park et al. [ 10] utilized Nighttime Light (NTL) inten...

work page 2021
[3]

A query is then performed on the tree by passing the grid coordinates as the region of interest, which returns all building polygons within that grid

tree. A query is then performed on the tree by passing the grid coordinates as the region of interest, which returns all building polygons within that grid. From this, we derive the total building count and the median building area of all included build- ings. In this framework, planned spacious areas are conventionally assigned higher scores relative to ...

work page
[4]

less -developed

To standardize these thresholds, we conducted a manual inspection of a large ran- dom sample of grids across nine representative cities to define the thresholds for the Red, Amber, and Green classes. Interestingly, our analysis revealed that Red grids consistently accounted for no more than 5% of the total urban area, while Amber grids never exceeded 9%. ...

work page
[5]

Mapping Urban Land Use at Street Block Level Using OpenStreetMap, Remote Sensing Data, and Spatial Metrics

Grippa T, Georganos S, Zarougui S, et al. Mapping Urban Land Use at Street Block Level Using OpenStreetMap, Remote Sensing Data, and Spatial Metrics. ISPRS International Journal of Geo-Information. 2018;7(7):246

work page 2018
[6]

Towards a scalable and transferable approach to map deprived areas using Sentinel-2 images and machine learning

Owusu M, Nair A, Jafari A, Thomson D, Kuffer M, Engstrom R. Towards a scalable and transferable approach to map deprived areas using Sentinel-2 images and machine learning. Comput Environ Urban Syst. 2021

work page 2021
[7]

Continental-Scale Building Detection from High Resolution Satellite Imagery

W. Sirko, S. Kashubin, M. Ritter, A. Annkah, Y.S.E. Bouchareb, Y. Dauphin, D. Keysers, M. Neumann, M. Cisse, J.A. Quinn. Continental-scale building detection from high resolu- tion satellite imagery. arXiv:2107.12283, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[8]

The Impact of Building Morphology on Energy Use Inten- sity of High -Rise Residential Clusters: A Case Study of Hangzhou, China

Feng W, Chen J, Yang Y, et al. The Impact of Building Morphology on Energy Use Inten- sity of High -Rise Residential Clusters: A Case Study of Hangzhou, China. Buildings. 2024;14(7):2245-2245

work page 2024
[9]

Incorporating Building Morphology Data to Improve Urban Land Use Mapping: A Case Study of Shenzhen

Zhang J, Song F, Wang Y, et al. Incorporating Building Morphology Data to Improve Urban Land Use Mapping: A Case Study of Shenzhen. Remote Sensing. 2025;17(16):2811

work page 2025
[10]

Monitoring urban land cover change: An expert system approach to land cover classification of semiarid to arid urban centers

Stefanov WL, Ramsey MS, Christensen PR. Monitoring urban land cover change: An expert system approach to land cover classification of semiarid to arid urban centers. Remote Sens- ing of Environment. 2001;77(2):173-185

work page 2001
[11]

XGBoost: a Scalable Tree Boosting System

Chen T, Guestrin C. XGBoost: a Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD ’16. 2016;1(1):785-794

work page 2016
[12]

Building(s and) cities: Delineating urban areas with a machine learning algorithm

Arribas-Bel D, Garcia-López MÀ, Viladecans-Marsal E. Building(s and) cities: Delineating urban areas with a machine learning algorithm. J Urban Econ. 2019;109:63-80

work page 2019
[13]

Liu, M., Li, R., Li, Y. et al. Nighttime light intensity and brightness suitability in urban functional zones. Sci Rep 15, 25113 (2025)

work page 2025
[14]

Learning Multidimensional Urban Poverty Representation with Satellite Imagery

Park S, Lee S, Kim J, et al. Learning Multidimensional Urban Poverty Representation with Satellite Imagery. arXiv.org. Published 2025. Accessed January 29, 2026

work page 2025
[15]

From chaos to clarity: Leveraging next-generation multimodal large language models for structured insights from unstructured data

Banerjee D, Thakur S, Bansal G, Babu SB, Varankar S, Khan MR. From chaos to clarity: Leveraging next-generation multimodal large language models for structured insights from unstructured data

work page
[16]

U -Net: Convolutional networks for biomedical image segmentation

Ronneberger O, Fischer P, Brox T. U -Net: Convolutional networks for biomedical image segmentation. arXiv (Cornell University). Published online May 18, 2015

work page 2015
[17]

Github.io

Wicket - Lightweight Javascript for WKT [Sandbox]. Github.io. Published 2026. Accessed January 22, 2026. https://arthur-e.github.io/Wicket/sandbox-gmaps3.html

work page 2026
[18]

Google LLC. (n.d.). Google Geocoding API. Google Maps Platform. Retrieved July 7, 2026, from https://developers.google.com/maps/documentation/geocoding

work page 2026
[19]

Binary codes capable of correcting deletions, insertions, and reversals

Levenshtein VI. Binary codes capable of correcting deletions, insertions, and reversals. So- viet Physics Doklady. 1966;10(8):707–710

work page 1966
[20]

STR: A simple and efficient algorithm for R - tree packing

Leutenegger ST, Edgington JM, Lopez MA . STR: A simple and efficient algorithm for R - tree packing. Proceedings of the 13th International Conference on Data Engineering . 1997:497–506

work page 1997
[21]

A density-based algorithm for discovering clusters in large spatial databases with noise

Ester M, Kriegel HP, Sander J, Xu X. A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD'96). AAAI Press; 1996:226-231

work page 1996
[22]

Virtues of the haversine

Sinnott RW. Virtues of the haversine. Sky Telesc. 1984;68(2):159

work page 1984
[23]

Libgeos.org

GEOS: geos::algorithm::hull::ConcaveHull Class Reference. Libgeos.org. Published 2026. Accessed March 18, 2026. https://libgeos.org/doxygen/classgeos_1_1algo- rithm_1_1hull_1_1ConcaveHull.html 34 S. Kulkarni and D. Banerjee

work page 2026
[24]

Credit Risk Scorecards

Siddiqi N. Credit Risk Scorecards. John Wiley & Sons; 2012

work page 2012
[25]

W., & Lemeshow, S

Hosmer, D. W., & Lemeshow, S. (2000). Applied Logistic Regression (2nd ed.). Wiley

work page 2000
[26]

Optuna - A hyperparameter optimization framework. Optuna. https://optuna.org/#code_ex- amples

work page
[27]

onlinemaps.surveyofindia.gov.in

Survey of India. onlinemaps.surveyofindia.gov.in. https://onlinemaps.survey- ofindia.gov.in/Digital_Product_Show.aspx

work page
[28]

High-Resolution Building and Road Detection from Sentinel-2

W. Sirko, E.A. Brempong, J.T.C. Marcos, A. Annkah, A. Korme, M.A. Hassen, K. Sapkota, T. Shekel, A. Diack, S. Nevo, J. Hickey, J.A. Quinn. High -Resolution Building and Road Detection from Sentinel-2. arXiv:2310.11622, 2023. Urban Neighborhood Delineation Using Open Buildings 35 INTERNAL 10 Appendix 10.1 Delinquency 59 Cities The cities are ordered based ...

work page internal anchor Pith review Pith/arXiv arXiv 2023

[1] [1]

developed a cost -effective approach to derive morphological features from open - source Sentinel-2 imagery, using XGBoost [7] models to identify informal settlements in Africa. Their findings reveal that these settlements possess heterogeneous local char- acteristics, supporting the hypothesis of varying lifestyle levels, and they developed a generalized...

work page

[2] [2]

AREA_BIN_

analyzed factors affecting nighttime lighting patterns, using brightness -based clus- tering to identify commercial and industrial zones. Their work established significant differences in brightness across various functional zones, including residential, under- developed, and industrial areas. Sungwon Park et al. [ 10] utilized Nighttime Light (NTL) inten...

work page 2021

[3] [3]

A query is then performed on the tree by passing the grid coordinates as the region of interest, which returns all building polygons within that grid

tree. A query is then performed on the tree by passing the grid coordinates as the region of interest, which returns all building polygons within that grid. From this, we derive the total building count and the median building area of all included build- ings. In this framework, planned spacious areas are conventionally assigned higher scores relative to ...

work page

[4] [4]

less -developed

To standardize these thresholds, we conducted a manual inspection of a large ran- dom sample of grids across nine representative cities to define the thresholds for the Red, Amber, and Green classes. Interestingly, our analysis revealed that Red grids consistently accounted for no more than 5% of the total urban area, while Amber grids never exceeded 9%. ...

work page

[5] [5]

Mapping Urban Land Use at Street Block Level Using OpenStreetMap, Remote Sensing Data, and Spatial Metrics

Grippa T, Georganos S, Zarougui S, et al. Mapping Urban Land Use at Street Block Level Using OpenStreetMap, Remote Sensing Data, and Spatial Metrics. ISPRS International Journal of Geo-Information. 2018;7(7):246

work page 2018

[6] [6]

Towards a scalable and transferable approach to map deprived areas using Sentinel-2 images and machine learning

Owusu M, Nair A, Jafari A, Thomson D, Kuffer M, Engstrom R. Towards a scalable and transferable approach to map deprived areas using Sentinel-2 images and machine learning. Comput Environ Urban Syst. 2021

work page 2021

[7] [7]

Continental-Scale Building Detection from High Resolution Satellite Imagery

W. Sirko, S. Kashubin, M. Ritter, A. Annkah, Y.S.E. Bouchareb, Y. Dauphin, D. Keysers, M. Neumann, M. Cisse, J.A. Quinn. Continental-scale building detection from high resolu- tion satellite imagery. arXiv:2107.12283, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021

[8] [8]

The Impact of Building Morphology on Energy Use Inten- sity of High -Rise Residential Clusters: A Case Study of Hangzhou, China

Feng W, Chen J, Yang Y, et al. The Impact of Building Morphology on Energy Use Inten- sity of High -Rise Residential Clusters: A Case Study of Hangzhou, China. Buildings. 2024;14(7):2245-2245

work page 2024

[9] [9]

Incorporating Building Morphology Data to Improve Urban Land Use Mapping: A Case Study of Shenzhen

Zhang J, Song F, Wang Y, et al. Incorporating Building Morphology Data to Improve Urban Land Use Mapping: A Case Study of Shenzhen. Remote Sensing. 2025;17(16):2811

work page 2025

[10] [10]

Monitoring urban land cover change: An expert system approach to land cover classification of semiarid to arid urban centers

Stefanov WL, Ramsey MS, Christensen PR. Monitoring urban land cover change: An expert system approach to land cover classification of semiarid to arid urban centers. Remote Sens- ing of Environment. 2001;77(2):173-185

work page 2001

[11] [11]

XGBoost: a Scalable Tree Boosting System

Chen T, Guestrin C. XGBoost: a Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining - KDD ’16. 2016;1(1):785-794

work page 2016

[12] [12]

Building(s and) cities: Delineating urban areas with a machine learning algorithm

Arribas-Bel D, Garcia-López MÀ, Viladecans-Marsal E. Building(s and) cities: Delineating urban areas with a machine learning algorithm. J Urban Econ. 2019;109:63-80

work page 2019

[13] [13]

Liu, M., Li, R., Li, Y. et al. Nighttime light intensity and brightness suitability in urban functional zones. Sci Rep 15, 25113 (2025)

work page 2025

[14] [14]

Learning Multidimensional Urban Poverty Representation with Satellite Imagery

Park S, Lee S, Kim J, et al. Learning Multidimensional Urban Poverty Representation with Satellite Imagery. arXiv.org. Published 2025. Accessed January 29, 2026

work page 2025

[15] [15]

From chaos to clarity: Leveraging next-generation multimodal large language models for structured insights from unstructured data

Banerjee D, Thakur S, Bansal G, Babu SB, Varankar S, Khan MR. From chaos to clarity: Leveraging next-generation multimodal large language models for structured insights from unstructured data

work page

[16] [16]

U -Net: Convolutional networks for biomedical image segmentation

Ronneberger O, Fischer P, Brox T. U -Net: Convolutional networks for biomedical image segmentation. arXiv (Cornell University). Published online May 18, 2015

work page 2015

[17] [17]

Github.io

Wicket - Lightweight Javascript for WKT [Sandbox]. Github.io. Published 2026. Accessed January 22, 2026. https://arthur-e.github.io/Wicket/sandbox-gmaps3.html

work page 2026

[18] [18]

Google LLC. (n.d.). Google Geocoding API. Google Maps Platform. Retrieved July 7, 2026, from https://developers.google.com/maps/documentation/geocoding

work page 2026

[19] [19]

Binary codes capable of correcting deletions, insertions, and reversals

Levenshtein VI. Binary codes capable of correcting deletions, insertions, and reversals. So- viet Physics Doklady. 1966;10(8):707–710

work page 1966

[20] [20]

STR: A simple and efficient algorithm for R - tree packing

Leutenegger ST, Edgington JM, Lopez MA . STR: A simple and efficient algorithm for R - tree packing. Proceedings of the 13th International Conference on Data Engineering . 1997:497–506

work page 1997

[21] [21]

A density-based algorithm for discovering clusters in large spatial databases with noise

Ester M, Kriegel HP, Sander J, Xu X. A density-based algorithm for discovering clusters in large spatial databases with noise. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD'96). AAAI Press; 1996:226-231

work page 1996

[22] [22]

Virtues of the haversine

Sinnott RW. Virtues of the haversine. Sky Telesc. 1984;68(2):159

work page 1984

[23] [23]

Libgeos.org

GEOS: geos::algorithm::hull::ConcaveHull Class Reference. Libgeos.org. Published 2026. Accessed March 18, 2026. https://libgeos.org/doxygen/classgeos_1_1algo- rithm_1_1hull_1_1ConcaveHull.html 34 S. Kulkarni and D. Banerjee

work page 2026

[24] [24]

Credit Risk Scorecards

Siddiqi N. Credit Risk Scorecards. John Wiley & Sons; 2012

work page 2012

[25] [25]

W., & Lemeshow, S

Hosmer, D. W., & Lemeshow, S. (2000). Applied Logistic Regression (2nd ed.). Wiley

work page 2000

[26] [26]

Optuna - A hyperparameter optimization framework. Optuna. https://optuna.org/#code_ex- amples

work page

[27] [27]

onlinemaps.surveyofindia.gov.in

Survey of India. onlinemaps.surveyofindia.gov.in. https://onlinemaps.survey- ofindia.gov.in/Digital_Product_Show.aspx

work page

[28] [28]

High-Resolution Building and Road Detection from Sentinel-2

W. Sirko, E.A. Brempong, J.T.C. Marcos, A. Annkah, A. Korme, M.A. Hassen, K. Sapkota, T. Shekel, A. Diack, S. Nevo, J. Hickey, J.A. Quinn. High -Resolution Building and Road Detection from Sentinel-2. arXiv:2310.11622, 2023. Urban Neighborhood Delineation Using Open Buildings 35 INTERNAL 10 Appendix 10.1 Delinquency 59 Cities The cities are ordered based ...

work page internal anchor Pith review Pith/arXiv arXiv 2023