pith. sign in

arxiv: 2604.22433 · v1 · submitted 2026-04-24 · 💻 cs.LG

Beyond Land Surface Temperature: Explainable Spatial Machine Learning Reveals Urban Morphology Effects on Human-Centric Heat Stress

Yuan Wang , Shengao Yi , Xiaojiang Li , Pengyuan Liu , Zhiwei Yang , Ronita Bardhan , Rudi Stouffs This is my paper

Pith reviewed 2026-05-08 12:21 UTC · model grok-4.3

classification 💻 cs.LG
keywords urban heat stressland surface temperatureuniversal thermal climate indexmachine learningspatial analysisurban morphologyexplainable AISingapore
0
0 comments X

The pith

Sky view factor drives most variation in human heat stress but contributes little to land surface temperature, showing LST misses key shading effects.

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

The paper compares 30-meter land surface temperature data from Landsat with 1-meter universal thermal climate index values in Singapore to test whether LST serves as a reliable proxy for actual human heat exposure. It applies geographically weighted XGBoost models and SHAP explanations to link both metrics to 2D and 3D urban features such as building density and sky openness. The analysis reveals clear differences in spatial patterns and shows that sky view factor strongly influences UTCI but has only marginal independent effect on LST. This indicates that LST overlooks the radiative and shading processes that govern real human comfort. The findings support shifting urban planning toward physiologically based heat metrics for more effective heat-risk management.

Core claim

Using a Modeling-Comparing-Assessing framework with Landsat LST and GPU-accelerated 1-m UTCI in Singapore, the study finds notable spatial discrepancies between the two thermal metrics and substantial non-stationary effects from urban morphology factors. Spatially explicit SHAP values show sky view factor as central to UTCI variability yet marginal for LST, while SHAP-GAM links higher albedo to increased UTCI. These results demonstrate that LST inadequately captures the shading-driven and radiative processes that govern human heat stress.

What carries the argument

Geographically weighted XGBoost (GW-XGBoost) combined with spatially explicit SHAP explanations applied to 2D and 3D urban morphology factors to compare LST and UTCI.

If this is right

  • Urban planners should incorporate physiologically relevant indices like UTCI rather than relying primarily on LST for heat-risk assessment.
  • Higher surface albedo may increase human heat stress despite cooling surface temperatures.
  • 3D urban features such as sky view factor require location-specific consideration because their effects on heat stress vary spatially.
  • Targeted interventions like improved shading can reduce human heat stress even where they have limited impact on satellite-measured LST.

Where Pith is reading between the lines

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

  • Satellite LST maps used for city-wide heat planning may systematically understate the benefits of adding shade structures in dense neighborhoods.
  • The approach could be extended to other cities by testing whether similar sky-view discrepancies appear when comparing LST and UTCI at matching resolutions.
  • Integration with pedestrian-level wind and humidity sensors would help isolate whether the sky view factor effect on UTCI is truly independent of airflow.
  • This distinction matters for equity, as areas with low sky view factor often coincide with dense, lower-income districts where heat exposure affects vulnerable populations most.

Load-bearing premise

The 1-meter UTCI model and chosen urban morphology variables fully capture human heat stress without major missing influences from wind, humidity, activity, or other confounders, and the identified relationships reflect causal mechanisms rather than mere correlations.

What would settle it

Direct field measurements of pedestrian thermal comfort or physiological responses across sites with differing sky view factors that align more closely with LST patterns than UTCI patterns, or that remove the sky view factor effect once local wind is accounted for.

Figures

Figures reproduced from arXiv: 2604.22433 by Pengyuan Liu, Ronita Bardhan, Rudi Stouffs, Shengao Yi, Xiaojiang Li, Yuan Wang, Zhiwei Yang.

Figure 1
Figure 1. Figure 1: Study area with 1-m land use/land cover (LULC), building height, and canopy height layers, and the workflow view at source ↗
Figure 2
Figure 2. Figure 2: Spatial distribution of 30-m LST (a) and 1-m UTCI (b), and the dynamic variations of UTCI at 8:00, 11:00, view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between LST and human-centric UTCI across LULC types. view at source ↗
Figure 4
Figure 4. Figure 4: Spatial relationship and quantitative divergence between surface temperatures and human-centric heat stress. view at source ↗
Figure 5
Figure 5. Figure 5: Global and local Moran’s I statistics for LST and UTCI. A hybrid Queen–nearest-neighbor spatial-weights view at source ↗
Figure 6
Figure 6. Figure 6: Spatially local R2 and standardized residuals of the GW-XGBoost model for LST (a-b) and UTCI (c-d) across subzones of Singapore. structural sensitivity underscores the potential for targeted urban design interventions to mitigate heat stress, a theme further explored through feature importance analysis in Section 3.3. 3.3. The difference of influencing factors in LST and UTCI revealed by explainable spatia… view at source ↗
Figure 7
Figure 7. Figure 7: SHAP global feature importance and local summary (beeswarm) plots for LST (a) and UTCI (b). view at source ↗
Figure 8
Figure 8. Figure 8: Spatial variation of the top-five feature importance ranks across subzones for (a) LST and (b) UTCI. Each view at source ↗
Figure 9
Figure 9. Figure 9: Spatially explicit results of the GW-XGBoost model for LST and UTCI across subzones of Singapore. (a-c) view at source ↗
Figure 10
Figure 10. Figure 10: SHAP dependence plots with GAM smoothers showing the non-linear relationships between the six most view at source ↗
Figure 11
Figure 11. Figure 11: SHAP dependence of Albedo with Sky View Factor (SVF) interaction. Scatter points indicate the SHAP view at source ↗
read the original abstract

Heat exposure connects the built environment and public health, directly shaping the livability and sustainability of urban areas. Understanding the spatial heterogeneity of heat exposure and its drivers is vital for climate-adaptive urban planning. However, most planning-oriented studies rely on land surface temperature (LST), and whether LST adequately represents human heat exposure and how it differs from physiologically relevant heat stress remains insufficiently examined. Here, adopting Landsat-retrieved 30-m LST and GPU-accelerated 1-m universal thermal climate index (UTCI) in Singapore, this study establishes a comprehensive "Modeling-Comparing-Assessing" framework to systematically evaluate the spatial and mechanistic discrepancies between the two metrics. We further investigate pronounced non-stationary and threshold-based quantitative relationships of the two metrics with urban factors by employing a novel geographically weighted XGBoost (GW-XGBoost) and generalized additive model (GAM) workflow. Our results demonstrate notable discrepancies in spatial patterns of LST and UTCI, along with substantial spatial heterogeneity in how 2D and 3D urban factors impact these two thermal metrics, as revealed by explainable GW-XGBoost models (global out-of-bag R2 = 0.855 for LST and 0.905 for UTCI, respectively). Crucially, spatially explicit SHAP interprets that sky view factor plays a central role in explaining UTCI variability but exhibits a comparatively marginal independent contribution to LST, indicating that LST inadequately captures shading-driven and radiative processes governing actual human heat stress. Notably, SHAP-GAM analysis indicates that higher albedo is associated with increased UTCI. These novel findings provide evidence for integrating physiologically relevant thermal indices to inform targeted heat risk management and climate-adaptive urban planning.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a 'Modeling-Comparing-Assessing' framework using 30-m Landsat LST and 1-m GPU-accelerated UTCI data for Singapore to compare land surface temperature with human-centric heat stress metrics. It employs geographically weighted XGBoost (GW-XGBoost) models with SHAP explanations and generalized additive models (GAM) to reveal spatial heterogeneity and non-stationary relationships with 2D/3D urban morphology factors, claiming that sky view factor (SVF) has a central role in UTCI but marginal contribution to LST, thus indicating LST inadequately captures shading and radiative processes relevant to human heat stress. Global out-of-bag R² values are reported as 0.855 for LST and 0.905 for UTCI.

Significance. If the central claim regarding the differential role of SVF holds after addressing potential resolution artifacts, this work would be significant for urban climate research and planning, as it provides evidence-based arguments for shifting from LST to physiologically relevant indices like UTCI in heat risk assessment, supported by high predictive performance and explainable spatial ML techniques. The integration of high-resolution modeling and interpretable ML is a strength.

major comments (2)
  1. [Results (SHAP interpretations)] The claim that sky view factor (SVF) has a central role in explaining UTCI variability but marginal independent contribution to LST (as per abstract and SHAP-GAM analysis) may be driven by the resolution mismatch between the 30-m LST and 1-m UTCI data. Since LST spatially averages fine-scale shading effects that SVF parameterizes at 1 m, the lower SHAP contribution to LST could be an aggregation artifact rather than evidence of LST's inadequacy in capturing shading-driven processes. An aggregation test or resolution-matched comparison is required to support this load-bearing interpretation.
  2. [Methods (model training and validation)] Details on cross-validation strategy, handling of spatial autocorrelation, error propagation, and sensitivity to data exclusions are missing despite the spatially explicit nature of the GW-XGBoost models. These are necessary to verify the reported out-of-bag R² values (0.855 for LST, 0.905 for UTCI) and the robustness of the non-stationary relationships identified.
minor comments (2)
  1. [Abstract] The abstract mentions 'notable discrepancies in spatial patterns' but does not quantify them; consider adding specific metrics or references to figures.
  2. [Discussion] The finding that higher albedo is associated with increased UTCI should be discussed in context of potential confounding factors such as material properties or measurement biases.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review, which highlights important considerations for strengthening the interpretation of our results and the transparency of our methods. We address each major comment below and outline the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: [Results (SHAP interpretations)] The claim that sky view factor (SVF) has a central role in explaining UTCI variability but marginal independent contribution to LST (as per abstract and SHAP-GAM analysis) may be driven by the resolution mismatch between the 30-m LST and 1-m UTCI data. Since LST spatially averages fine-scale shading effects that SVF parameterizes at 1 m, the lower SHAP contribution to LST could be an aggregation artifact rather than evidence of LST's inadequacy in capturing shading-driven processes. An aggregation test or resolution-matched comparison is required to support this load-bearing interpretation.

    Authors: We agree that the resolution difference between the 30-m LST and 1-m UTCI datasets is a relevant factor that merits explicit examination to rule out aggregation artifacts. Our core claim is that LST, as commonly available at 30 m, does not resolve the fine-scale shading and radiative processes captured by 1-m UTCI, which is why SVF shows lower independent contribution to LST. This is physically expected rather than purely artifactual, but we acknowledge that a direct test would provide stronger evidence. In the revised manuscript we will add an aggregation analysis: we will upscale the 1-m UTCI to 30-m resolution, retrain the GW-XGBoost model on the aggregated data, and compare the resulting SHAP values for SVF against the original 1-m results. This will be presented in a new supplementary figure and discussed in the Results section to confirm that the differential role of SVF persists beyond simple aggregation effects. revision: yes

  2. Referee: [Methods (model training and validation)] Details on cross-validation strategy, handling of spatial autocorrelation, error propagation, and sensitivity to data exclusions are missing despite the spatially explicit nature of the GW-XGBoost models. These are necessary to verify the reported out-of-bag R² values (0.855 for LST, 0.905 for UTCI) and the robustness of the non-stationary relationships identified.

    Authors: We appreciate this request for greater methodological transparency. The current manuscript reports global out-of-bag R² values from the GW-XGBoost models but does not fully detail the spatial cross-validation procedure, autocorrelation handling, or sensitivity checks. In the revised version we will expand the Methods section to include: (1) a description of the spatial blocking cross-validation strategy used to mitigate autocorrelation (with block sizes chosen based on variogram analysis); (2) explicit discussion of how input data uncertainties (e.g., from Landsat LST retrieval and UTCI modeling) propagate through the models; and (3) results from sensitivity tests that systematically exclude subsets of data (e.g., by land-use type or spatial region) to assess stability of the reported R² values and the identified non-stationary relationships. These additions will be supported by new supplementary tables and text. revision: yes

Circularity Check

0 steps flagged

No significant circularity; data-driven ML comparison is self-contained

full rationale

The paper trains separate GW-XGBoost models on urban morphology predictors to explain 30 m LST and 1 m UTCI, then applies SHAP for interpretation. These are standard supervised fits on independent inputs (satellite retrievals and GPU-modeled index) with no equations that reduce outputs to inputs by construction, no self-citations invoked as uniqueness theorems, and no ansatz or renaming that collapses the claimed discrepancy. Global OOB R² values and SHAP rankings are direct model outputs, not tautological. Minor hyperparameter choices exist but do not load-bear the central LST-vs-UTCI mechanistic claim.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Only abstract available so ledger is partial; relies on standard remote sensing and ML assumptions plus domain-specific thermal modeling choices.

axioms (2)
  • domain assumption The 1-m resolution GPU-accelerated UTCI model accurately represents physiologically relevant human heat stress
    Invoked to justify using UTCI as the human-centric ground truth against which LST is compared
  • domain assumption Selected 2D and 3D urban morphology factors are the primary drivers of spatial variation in both thermal metrics
    Basis for the GW-XGBoost feature set and subsequent SHAP interpretations

pith-pipeline@v0.9.0 · 5632 in / 1489 out tokens · 44697 ms · 2026-05-08T12:21:54.708194+00:00 · methodology

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

2 extracted references · 2 canonical work pages

  1. [1]

    Building and Environment 288, 113992

    Improving WRF-BEP+BEM performance in simulating wind-temperature-humidity in high-density urban areas: A case study of a megacity. Building and Environment 288, 113992. doi:10.1016/j.enbuild.2025.113992. Huang, B., Wu, B., Barry, M., 2010. Geographically and temporally weighted regression for modeling spatio-temporal variation in house prices. Internation...

    work page doi:10.1016/j.enbuild.2025.113992 2025

  2. [2]

    furnace city

    Global urban population exposure to extreme heat. Proceedings of the National Academy of Sciences 118, e2024792118. U.S. Geological Survey, 2025. Landsat 8 mission — u.s. geological survey. URL:https://www.usgs. gov/landsat-missions/landsat-8. Wan, Y., Du, H., Yuan, L., Xu, X., Tang, H., Zhang, J., 2025. Exploring the influence of block environmental char...

    work page arXiv 2025