Shanghai study links land cover with relative summer surface heat

Built-up areas in Shanghai consistently occupied the warmer end of the city’s land surface temperature distribution, while water bodies remained markedly cooler across four Landsat observations from 2000 to 2024, according to a remote-sensing study that combined land-cover stratification, explainable machine learning and spatial validation. Vegetation generally occupied a cooler relative position but showed substantial variation, indicating that the thermal signature of an environmental indicator differed across land-cover contexts. In the fitted model, the normalized difference built-up index, NDBI, made the largest SHAP-based contribution to predictions of relative thermal conditions, ahead of the normalized difference water index, NDWI, and normalized difference vegetation index, NDVI. These results describe associations within the selected images and grid cells, rather than establishing independent causal effects or direct measures of human heat exposure.

The study examined summer acquisitions dated August 1, 2000, August 5, 2010, August 16, 2020, and August 3, 2024, using Landsat Collection 2 Level-2 products. These were discrete observations rather than a continuous annual series, so the analysis compares selected spatial patterns across years rather than tracking continuous change. Land surface temperature, NDVI, NDBI and NDWI were aggregated from 30-metre satellite pixels into regular 300-by-300-metre grid cells. Each cell was assigned one of four dominant land-use and land-cover categories: vegetation, built-up land, water, or “other,” a mixed class containing bare land and marshes. Land-cover classification used supervised maximum-likelihood classification, with overall accuracy ranging from 93.11% in 2000 to 97.68% in 2024 and Kappa coefficients from 0.87 to 0.96.

To compare observations taken under different city-wide thermal backgrounds, the researchers converted land surface temperature in each grid cell into a thermal anomaly, or TA, by subtracting the mean land surface temperature for that acquisition date and dividing by its standard deviation. A positive TA therefore means that a cell was warmer than the city-wide mean on that date, while a negative value means it was cooler. Because this is a linear within-date standardization, it preserves the ranking of cells and creates no new ordering information; LST and TA consequently produced identical statistical comparisons. The measure is also not a climatological anomaly, an estimate of air temperature, or a direct measure of heat risk. Its purpose was to describe each cell’s relative thermal position within a particular image.

That relative ranking revealed pronounced differences among land-cover types. Built-up cells had the highest median land surface temperature in every selected year, reaching 42.04 °C in 2000, 42.95 °C in 2010, 45.75 °C in 2020 and 50.00 °C in 2024. Median temperatures over water were markedly lower at 29.92 °C in 2000, 29.99 °C in 2010, 31.74 °C in 2020 and 32.88 °C in 2024. Median TA values for built-up areas ranged from 0.98 to 1.34, compared with negative values from −2.00 to −1.30 for water. Vegetation generally fell below the date-specific mean, but some vegetated cells still occupied relatively warm positions. The “other” category had intermediate and more variable characteristics, consistent with its mixture of bare surfaces, marsh margins and transitional environments.

Non-parametric tests confirmed that the four land-cover groups differed for all five thermal and spectral indicators in every year, with Kruskal–Wallis tests yielding p-values below 0.001. Epsilon-squared effect sizes, which ranged from 0.448 to 0.719, indicated large differences. Dunn–Bonferroni post-hoc comparisons also found every pairwise land-cover difference significant after adjustment for multiple testing. The land-cover composition itself changed over the selected observations: vegetation declined from 61.8% of valid grid cells in 2000 to 49.7% in 2010 and about 50% in 2020 and 2024, while built-up land increased from 19.2% to 32.8% and then remained near 32.5%. Water’s share declined from 12.8% to 9.5%. These contrasts across dates are descriptive and do not by themselves demonstrate a continuous long-term trajectory.

The researchers then fitted a single XGBoost regression model to the pooled TA observations, using NDVI, NDBI and NDWI as predictors. Land-cover categories were deliberately excluded from the model itself and instead used to stratify and interpret the SHAP responses. The model used 400 estimators, a learning rate of 0.05, a maximum depth of six, a minimum child weight of 5, a subsample ratio of 0.80, and L1 and L2 regularization parameters of 0.10 and 1.00. NDBI had the largest mean absolute SHAP value, 0.48, compared with 0.19 for NDWI and 0.14 for NDVI. Higher NDBI was mainly associated with positive SHAP contributions, indicating higher predicted relative temperatures; higher NDWI was predominantly associated with negative contributions, while NDVI generally shifted predictions downward but with a more variable response pattern. These rankings describe contributions within the fitted model, not independent causal influence.

Land-cover stratification added detail to those overall attributions. In vegetation and built-up cells, increasing NDVI shifted SHAP values toward more negative contributions, associating greater vegetation with lower predicted TA. Water and the mixed “other” category displayed more complex, non-monotonic NDVI responses. NDBI showed clear increases in SHAP response for vegetation and built-up cells, with the strongest and fastest positive response in built-up areas. Responses for water and “other” land were more nonlinear, reflecting their broader and more transitional surface mixtures. NDWI was associated with declining SHAP values in all four categories. Water had one of the strongest negative responses, reaching a relatively stable negative range at high NDWI values, whereas vegetation and built-up cells showed smoother declines. Because the indices have overlapping spectral bands, these patterns are descriptive associations rather than formal tests of interactions between land cover and spectral indicators.

Spatial dependence was unusually strong: global Moran’s I values for TA ranged from 0.882 to 0.905 across the four dates, with permutation-based p-values of 0.001. To test whether conventional random validation inflated performance, the team used five-fold spatial block cross-validation with 6-by-6-kilometre blocks, keeping observations from the same location in the same fold. Performance declined only slightly under this stricter scheme. The coefficient of determination fell from 0.777 under random cross-validation to 0.769 under spatial block validation, while root mean square error rose from 0.472 to 0.478 and mean absolute error from 0.346 to 0.351. The limited reduction suggests that predictive performance was comparatively stable under the selected partitioning scheme, while also showing why spatial validation is a more conservative assessment of geographic transferability for strongly autocorrelated data.

The findings support preliminary thermal screening by land-cover type, but they do not constitute a comprehensive model of urban heat or evidence that any specific intervention would reduce temperatures. The model omitted potentially important influences including building morphology, landscape configuration, meteorology and socioeconomic exposure, and it did not reconstruct land-cover transitions or population vulnerability. The four selected summer images may also differ in acquisition conditions, sensor characteristics and background climate, even after standardization. In addition, the 300-metre aggregation and dominant-category assignment limit the resolution of the analysis, and the mixed “other” category combines surfaces with potentially different thermal behaviour. The combined use of land-cover comparison, spatial validation and SHAP interpretation nevertheless offers a transparent way to identify where built surfaces, vegetation and water occupy different relative thermal positions, providing a basis for more detailed urban heat assessment rather than a direct prescription for adaptation.

Subject of Research: Land-cover-specific associations between spectral environmental indicators and relative surface thermal conditions in Shanghai

Article Title: LULC-stratified associations between surface environmental indicators and standardized land surface temperature: an XGBoost–SHAP analysis

Article References: Zhang, T., & Xu, R. (2026). LULC-stratified associations between surface environmental indicators and standardized land surface temperature: an XGBoost–SHAP analysis. Frontiers in Earth Science, 14, Article 1902432. https://doi.org/10.3389/feart.2026.1902432

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1902432

Keywords: Shanghai, land surface temperature, urban heat, land cover, remote sensing, XGBoost, SHAP, spatial validation, NDBI, NDWI, NDVI, LULC-stratified