Vernacular Settlement Fabric and Within-Village Surface-Temperature Heterogeneity in Huizhou Traditional Villages: An Architecture-Informed Multiscale Analysis

The following sections specify the regional frame, settlement-fabric measures, thermal outcome and village-held-out validation procedures used to operationalize the sequence outlined above.

2.1 Study region, village frame and village-date sample

The regional analytical frame contains 60 traditional villages across six county areas of historic Huizhou: She, Yi, Xiuning, Qimen, Jixi and Wuyuan (Figure 1). Wuyuan lies in present-day Jiangxi Province and the other five county areas in Anhui Province. The frame is treated as a purposive regional comparison frame rather than a probability sample.

The primary thermal sample was defined by Landsat availability, the quality criteria in Section 2.3, complete predictor availability and a minimum of ten valid analysis units per village-date event. It contains 1,121 analysis-unit observations nested within 33 village-date events, covering 832 unique analysis units in 25 villages.

The three qualifying acquisition dates were 24 May 2019, 28 May 2024 and 29 May 2024, contributing 663, 102 and 356 analysis-unit observations, respectively. These dates were retained because they satisfied the same Landsat availability and quality rules; their close calendar alignment within 24–29 May also provides a common late-spring seasonal frame for cross-year comparison. Two values outside the fixed 0–70 ℃ QC range were excluded before modeling. Each retained event contains at least ten analysis units satisfying both predictor and thermal criteria, and minimum valid-unit requirements from 5 to 30 units were evaluated in sensitivity analyses.

Selection balance between the 25-village thermal sample and the remaining 35 framed villages was assessed using standardized mean differences across morphology and terrain variables (Figure S1). The complementary outcome-blind weighting check used pre-outcome morphology and terrain information from the regional frame, excluding ΔLST from weight construction, and was applied only as a fixed-effect sensitivity; its directional result is reported in Section 3.6. Table 2 summarizes the study design, common response geometry, morphology contexts and analytical roles.

Table 2. Study design, data sources and analytical roles

2.2 Architectural settlement-fabric representation and terrain embedding

The settlement-fabric representation uses four linked levels: metrics are computed variables, constructs group those variables architecturally, the Huizhou spatial reading interprets the constructs at regional scale, and local evidence identifies what is needed for site-scale diagnosis. Five architectural construct groups were used: plan-area occupation, building-plan grain, spacing and locational regularity, continuity and interstitial openness, and directional order. Building metrics were derived from GlobalBuildingAtlas (GBA), an open product of building polygons and heights [21]. Plan-area occupation is quantified by building coverage ratio (BCR), while Clark–Evans R follows the nearest-neighbor regularity formulation of Clark and Evans [22]; the remaining metrics describe object grain, spacing, aggregation/open gaps and orientation.

GBA is based mainly on 2019 PlanetScope imagery, with 2018 imagery used where necessary, so the morphology layer is treated as a static circa-2018/2019 representation for both thermal years. Slope and topographic position index (TPI) from FABDEM [23] describe terrain embedding.

GBA median height, height interquartile range, height coefficient of variation and a simple volume-density proxy provide neighborhood-scale vertical context. They were tested separately from the primary 2D morphology set so that vertical context could be evaluated without changing the primary comparison among terrain, settlement morphology and their combination. Table 1 shows the principal vertical-context descriptors used for architectural reading; the complete GBA-derived vertical-context sensitivity, based on the 3D building-height product, is reported in Table S1.

Figure 3 provides an illustrative architectural plan-reading reference for interpreting the settlement-fabric constructs in Table 1 at plan scale. Fuxi Village is included in the thermal sample, and the figure-ground reading is derived from the same regional building/morphology database used to construct the settlement-fabric predictors.

3.png

Figure 3. Fuxi Village figure-ground and plan-grain reading derived from the regional building/morphology database. Panel (a) shows the whole-village solid–void pattern; dashed detail windows B and C correspond to panels (b) and (c), which highlight plan-grain and solid–void relations

Table 1 maps these metrics to their architectural reading and evidence boundary. The construct framework follows established urban-morphology concepts of tissue, grain, plan form and morphological structure [11-13].

Within this framework, high plan-area occupation (BCR) can coexist with narrow lanes and internal tianjing voids; open gaps between building masses may contain lanes, courtyards, planting, water edges or residual space; and directional order may reflect routes, slope or hydrological structure. The regional measures therefore identify spatial conditions for closer architectural diagnosis while local evidence resolves their detailed spatial meaning.

Across whole-village and detail views, the Fuxi figure-ground plate links solid–void pattern, plan grain, adjacency and interstitial space to the regional construct vocabulary.

The georeferenced 100 × 100 m analysis unit was retained as the common response geometry. Settlement morphology was summarized around the same unit at 100, 300 and 500 m; these distances define the surrounding morphology context and do not create new thermal pixels. The 300 m context was used for the primary models, while 100 and 500 m were retained for the common-sample scale comparison. Figure 4 uses Fuxi Village, Yucun Township, Xiuning County, to show the same analysis units embedded in the three morphology contexts.

Fuxi Village was selected because it provides two-date 2019/2024 coverage and broad analysis-unit coverage, allowing the multiscale geometry to be shown clearly. Figure 4 illustrates how the same response geometry can be read within progressively broader settlement contexts without changing the response unit itself.

4.png

Figure 4. Fuxi Village, Yucun Township, Xiuning County: analysis units and multiscale morphology context. Panel (a) shows the 100 × 100 m core, edge and context analysis units; panels (b–d) map GBA-derived building coverage ratio (BCR) summarized at 100, 300 and 500 m and assigned to the same analysis units. The context radius denotes the neighborhood used to summarize settlement fabric and does not create a coarser thermal response unit. These panels present the analytical morphology representation used for regional comparison

Core/edge/context labels describe each analysis unit relative to the GBA-derived settlement extent; they are spatial descriptors rather than cadastral, administrative or statutory heritage boundaries.

The modeled regional covariates comprise settlement morphology, terrain and the tested neighborhood-scale vertical context; independent vegetation, water and impervious/open-surface fractions remain part of the finer local evidence layer.

2.3 Landsat-derived surface temperature and quality control

Thermal estimates came from Landsat 8–9 Collection 2 Level-2 Surface Temperature (ST) products [24, 25]. The distributed product is provided on a 30 m grid, while the TIRS thermal bands have approximately 100 m native sampling. The 100 × 100 m analysis units are therefore georeferenced aggregation polygons over the ST product rather than independent sensor pixels. The operational LST retrieval is described by Malakar et al. [26] within the broader Landsat 8 science framework [27].

Source-pixel ST values were intersected exactly with each analysis unit and area-weighted. Valid observations required clear QA_PIXEL bits 0–5, ST_QA ≤ 2.0 K, cloud distance ≥ 1 km and at least 50% valid area. Scaled ST integers were converted using ST(K) = DN × 0.00341802 + 149.0 and then to ℃ following USGS guidance [25]. Each retained thermal estimate was linked to the same georeferenced analysis-unit geometry used throughout the study.

Because the analysis units are aggregation geometry over the thermal product, grouped validation, Moran/Conley diagnostics and a separate response-aggregation test address spatial dependence and scale sensitivity. Quality sensitivities applied a 75% valid-area rule, a 2 km cloud-distance rule and minimum valid-unit requirements of 5, 10, 15, 20 and 30 units per village-date event; the corresponding results are reported in Table 3.

Table 3. Thermal-quality and minimum-valid-unit sensitivity analyses, including the 2 km cloud-distance check

2.4 Within-village-date thermal outcome and event-equal weighting

For analysis unit i in village v on date t, the outcome was defined as:

ΔLST(i,v,t) = LST(i,v,t) − median[LST(v,t)]

Centering within each village-date event converts the response into a relative thermal pattern: positive values identify locations warmer than the contemporaneous village median and negative values identify cooler locations. The comparison therefore focuses on spatial differences inside the same village under the same Landsat overpass rather than on absolute temperature differences between villages or years.

Because events contained different numbers of valid units, each observation received weight 1/n(v,t), giving every village-date event equal total weight in model fitting and primary performance metrics.

2.5 Predictive models and grouped validation

This stage tests whether settlement-fabric relationships generalize to villages not used for model fitting. Ridge regression was the primary transfer model because it handles correlated predictors while retaining a transparent linear structure [28]. Predictors were standardized within each training fold; α = 1.0 was used for the primary model, with α = 0.1 and 10 as sensitivity checks. Random Forest regression [29] provided a nonlinear comparison, with implementation settings reported in the APPENDIX.

Primary transfer performance used five-fold village-grouped cross-validation (implemented with GroupKFold), holding out whole villages so that all analysis units from a village remained in the same fold. R², RMSE and MAE were calculated with event-equal weights. Alternative tests used geographic village folds, leave-three-county combinations, joint date-and-county holdouts and leave-one-Landsat-scene-out validation; an unweighted repetition tested the nested-model ordering [19, 20]. Local-coordinate controls used the planar centroid coordinates of the georeferenced analysis units as simple positional covariates; they were evaluated with terrain and with the full morphology + terrain transfer set under the same village folds to test whether the fabric constructs retained predictive value beyond simple spatial position; the comparison is reported with the predictor-set results in Section 3.4.

2.6 Panel inference, multiplicity and spatial dependence

This stage asks which associations remain stable after accounting for clustered and spatially correlated uncertainty. Predictive transfer and predictor-level inference use related but distinct variable sets. The transfer benchmark uses nine morphology predictors—plan-area occupation (BCR), median footprint area, mean compactness, mean nearest-neighbor distance, Clark–Evans R, aggregation index, 20 m porosity, orientation coherence and orientation entropy—plus elevation, slope and TPI, giving 12 predictors for morphology + terrain. The inferential specification uses 11 predictors. Outcome-blind prefit collinearity screening used variance inflation factors (VIFs); after mean nearest-neighbor distance was omitted, the retained 11-predictor set remained below the pre-specified VIF threshold. The remaining eight morphology predictors plus elevation, slope and TPI are reported in Tables S2a–b.

Model specification. Event-equal weighted least squares used the 11-predictor inferential set with village and acquisition-date fixed effects. Because morphology and terrain predictors are time-invariant at unit level, inference uses village and acquisition-date fixed effects rather than analysis-unit fixed effects. Village×date fixed effects and a centered plan-area occupation (BCR) × median-height interaction provided additional checks on event alignment and neighborhood-scale vertical context.

Uncertainty and robustness. Uncertainty was evaluated with village-clustered standard errors, Benjamini–Hochberg false-discovery-rate control [30], 50,000-draw village-level wild-cluster tests (a finite-sample cluster-robust check) and Conley spatial-HAC covariance (a distance-based covariance correction). Residual spatial dependence was evaluated with Moran’s I [31]. Conley spatial-HAC covariance [32] used analysis-unit centroid distances in EPSG:32650 with a Bartlett kernel at 200 m, 500 m and 1 km within village-date events; a same-date sensitivity extended the distance range to 1, 5 and 10 km across villages. These alternatives change the uncertainty model while leaving fitted coefficient estimates unchanged; complete predictor-level results are reported in Tables S2a–b.

2.7 Scale and complementary robustness analyses

Scale comparison. To test whether the regional relationship depends on morphology context, transfer was compared at 100, 300 and 500 m using analysis-unit observations complete at all three contexts. Re-applying the minimum 10-valid-unit requirement produced a common sample of 274 observations from eight villages and ten village-date events. The same village-grouped model was used at all three contexts, with village-block bootstrap uncertainty and a representativeness comparison against the other primary villages (Figure S2; Tables S3a–b).

Complementary morphology checks. Model-implied q10-to-q90 conditional contrasts summarize the difference in ΔLST between low (10th-percentile) and high (90th-percentile) predictor values while the fitted model holds other terms fixed; coefficient uncertainty is evaluated with Conley and wild-cluster procedures. Linear and spline plan-area occupation (BCR) responses were compared to evaluate response form and the presence of a strong threshold.

Additional robustness checks. Complementary checks evaluate selection weighting (Section 3.6), GBA-derived vertical-context sensitivity (Table S1), minimum valid-unit and thermal-quality sensitivity (Table 3), and response aggregation (Table S4). Section 3.6 examines plan-area occupation (BCR) response form and neighborhood-scale vertical context alongside the primary inferential analysis.

A separate response-aggregation test combined the quality-screened 100 m summaries into georeferenced 200 m blocks. LST was re-aggregated with the original valid-area weights and re-centered by village-date; morphology and terrain predictors were averaged across contributing 100 m analysis units. The test retained the original village folds and event-equal weighting and compared blocks with at least two, at least three, and all four contributing 100 m analysis units (Table S4). This isolates a change in response aggregation from the 100/300/500 m morphology-context comparison, which keeps the 100 m response geometry fixed.

Analytical scope. The study focuses on relative surface-temperature heterogeneity within the three qualifying late-May acquisitions in 2019 and 2024. Model output is interpreted as regional settlement-fabric screening that identifies locations and fabric conditions for finer architectural diagnosis.

The discussion follows the architectural evidence sequence: repeatability and transfer establish the value of regional comparison, construct and scale results show how settlement fabric can be read regionally, and robust associations are translated into local architectural diagnosis.

4.1 Repeatability and transfer establish a regional screening signal

For RQ1, cross-year repeatability and cross-village generalization provide complementary evidence for regional comparison. Matched late-May locations preserve their warm–cool ordering across 2019 and 2024, and overall out-of-sample performance remains positive under each validation scheme. Most constituent holdouts are positive; the few below-zero cases occur under the strongest county/date separation and identify where additional local calibration can refine the regional reading.

The value lies in the comparison itself. Centering temperature within each village focuses attention on spatial differences inside a settlement, and holding out entire villages asks whether settlement-fabric relationships learned elsewhere retain out-of-sample value in a new one. This pairing separates within-village thermal patterning from cross-village generalization.

4.2 Plan-area occupation leads the settlement-fabric hierarchy

For the construct component of RQ2, the hierarchy is architecturally more informative than a flat predictor list. Plan-area occupation leads the regional transfer hierarchy, while the spacing and locational regularity construct provides the second strongest construct-level signal. These two constructs distinguish occupation from configuration: one reads how strongly building mass occupies the plan, the other how mapped building objects are arranged. Similar occupation levels can contain different lane, tianjing, planted or water-linked gaps and surface conditions, so plan-area occupation (BCR) is read as a screening descriptor rather than a direct design prescription.

Site-scale evidence for lane and courtyard geometry, vegetation, water edges, materials and heritage fabric resolves the local spatial meaning of the regional descriptors [5, 6]. Building-plan grain, continuity/interstitial openness and directional order remain complementary descriptors for that local plan reading.

Regional surface temperature and pedestrian microclimate operate at different scales. Huizhou and related traditional-settlement research resolves height-to-width ratio (H/W), orientation, interface opening, water/green ratios and tianjing form [33-37]. GBA median height and height variability provide only neighborhood-scale vertical context; detailed street-canyon geometry, sky-view factor (SVF) and frontage continuity belong to the site-scale architectural diagnostic layer.

4.3 The 300 m result is a settlement context, not a design radius

For the scale component of RQ2, the analysis is best read as a settlement-context result. In the restricted common sample, 300 m provides the strongest performance among the tested morphology contexts. Architecturally, that context can span several adjoining building groups, lanes and open-space sequences, allowing occupation, spacing and directional order to be read as a local piece of settlement fabric around the 100 m response geometry. Figure 4 makes the distinction visible: the morphology context changes, while the response geometry remains the same. The 300 m result therefore describes settlement context rather than a courtyard module, street-section threshold or conservation radius.

The complementary 200 m aggregation sensitivity reinforces the same distinction: morphology-based transfer crosses below zero under the stricter ≥3 and complete four-unit definitions as sample coverage contracts. Context scale and response aggregation therefore answer different questions—the 300 m result concerns the neighborhood from which morphology is summarized, whereas the 100 m analysis unit is the common response alignment used in the primary analysis.

4.4 Robust associations sharpen the architectural reading

For RQ3, the stable predictor-level directions sharpen the architectural reading. Plan-area occupation (BCR) and locational regularity (Clark–Evans R) provide the consistently stable morphology directions, while slope and terrain position form the terrain-embedding context. Read together, they frame regional surface-temperature heterogeneity as a joint fabric–terrain condition rather than a single plan-occupation measure.

Architecturally, slope and TPI locate settlement fabric within hillside, valley and local-relief settings, broadening screening from plan occupation/configuration to terrain embedding. These regional associations are best carried forward as terrain-embedding cues; site-scale street–slope sections, drainage relationships and measured building–water interfaces provide the evidence needed for mechanism-specific interpretation.

4.5 Architecture-informed screening and heritage-compatible local diagnosis

The architecture and conservation translation is organized as a proposed five-stage diagnostic pathway (Figure 12). Regional screening identifies Landsat-derived surface-temperature patterns, while site-scale evidence is required to evaluate pedestrian microclimate and design performance. Georeferenced 100 × 100 m analysis units remain embedded in explicit multiscale settlement contexts, and the construct vocabulary established in Table 1 keeps the regional model legible in architectural terms.

The Fuxi Village figure-ground reference (Figure 3) and the Fuxi analytical views (Figures 4–5) bridge regional metrics and plan reading: the former shows solid–void grain and adjacency, while the latter locates the response geometry and morphology contexts used in the regional analysis.

12.png

Figure 12. Conservation-oriented five-stage diagnostic pathway. Regional analysis provides the screening layer; site-scale architectural evidence provides the local diagnostic layer. Heritage-value review precedes consideration of reversible intervention options, and post-intervention monitoring forms a distinct evidence stage. The pathway is conceptual, with site-specific evidence guiding any local intervention assessment

Within this pathway, Stage 1 uses regional screening to identify relative warm/cool areas together with their settlement-fabric context. Stage 2 adds local figure-ground, street, courtyard/open-space, water, vegetation, material, terrain and vertical-enclosure evidence. Stage 3 evaluates heritage value and character-defining fabric. Stage 4 considers reversible or low-impact options grounded in that diagnosis, and Stage 5 monitors both thermal and conservation outcomes. This sequence follows conservation principles of significance, authenticity, cautious change and minimum necessary intervention and keeps architectural and conservation judgment between regional screening and any site-specific action. Table 6 summarizes these evidence needs; the checklist organizes evidence requirements for site-specific assessment, with local heritage and performance review governing any intervention decision. Stages 3–5 function as conceptual decision stages; their use depends on site-specific heritage and performance evidence rather than intervention-effect estimates from the regional analysis.

Table 6 groups six local conditions with the corresponding architectural, environmental and heritage evidence. The regional BCR association helps prioritize inspection, while site-scale evidence determines how those conditions are interpreted in conservation practice.

Table 6. Diagnostic evidence checklist for conservation screening and local architectural assessment. The entries define evidence needs for site-specific assessment; heritage significance and locally established environmental effects guide any subsequent action. The review-intensity terms are indicative categories within this conceptual workflow, not statutory approval classifications

4.6 Limitations

Five limitations define the transfer range of the findings.

Surface composition. Independent vegetation, water and impervious/open-surface fractions form a finer evidence layer than the present regional covariate set, so plan-form variables may also carry surface-composition information.

Thermal observation scale. Landsat ST represents overpass-time surface temperature, and the 100 × 100 m analysis units are aggregation geometry over the resampled thermal product rather than a stand-alone physical thermal scale.

Temporal coverage. Thermal observations are limited to only three late-May acquisition days (24 May 2019 and 28–29 May 2024). The cross-year result therefore demonstrates repeatability only within this closely matched late-spring window and cannot establish that the same thermal distribution persists through summer or across other seasons. Multi-season observations are required before making seasonal or summer-wide generalizations.

Spatial dependence. Residual spatial autocorrelation remains after fixed effects and local spatial trends; complementary uncertainty procedures account for that structure in the regional inference.

Local architectural evidence. Site-scale architectural evidence forms the next diagnostic layer. The Fuxi database-derived figure-ground plate provides a plan reading of the same regional building source, while finer street, courtyard, material and water-edge resolution requires georeferenced local evidence. These boundaries position the study as regional architecture-informed screening that prioritizes locations and fabric conditions for site-scale diagnosis.

This paper was supported by the 2026 Hefei Municipal Philosophy and Social Science Planning Project (Grant No.: HFSKYY202609), the Jiaxing Science and Technology Program (Public Welfare Research Special Project) (Grant No.: 2025CGZ077), and the Anhui Jianzhu University Scientific Research Project for Introducing Talents (Grant No.: 2023QDZ21).