Introduction

Tropical Africa has emerged as a global hotspot of cropland expansion over the past two decades, driven by rising population and increasing demands for food, biofuels, and other agricultural products1,2,3,4. The conversion of natural biomes into croplands for crop cultivation alters greenhouse gas (GHG) fluxes, thereby contributing to global climate change5,6,7,8. In addition, locally, cropland conversion can directly modify the land surface energy budget and thereby affect land surface temperature (Ts) through changes in key biophysical properties, including surface albedo (α), canopy structure, and evapotranspiration9,10,11,12,13, and further influence regional crop growth and agricultural productivity14,15,16. The magnitude and direction of these effects depend on the background natural biomes of the converted land9,11,17,18 and can be further modulated by land management practices, such as irrigation19,20,21, which in some cases may be comparable in magnitude to the biogeochemical effects associated with GHG flux perturbations22,23.

The biophysical impacts of croplands on Ts vary diurnally owing to differences in the underlying mechanisms11,24, with one notable distinction being that α influences the land surface energy balance only in the presence of sunlight25,26. Although previous studies have investigated the daytime and nighttime biophysical effects of croplands on Ts using Moderate Resolution Imaging Spectroradiometer (MODIS) observations11,24, the restricted temporal resolution and absence of diurnal measurements of surface biophysical properties have constrained mechanistic insights. Consequently, studies incorporating full diurnal observations of all relevant contributing variables are needed. Moreover, past investigations have often overlooked the role of background hydroclimatic conditions in modulating cropland biophysical impacts. For instance, in arid regions, crop cultivation typically relies on irrigation and other anthropogenic interventions, which can substantially modify the surface properties of croplands relative to the native biomes they replace27,28,29. Accounting for the hydroclimatic context is therefore essential for mechanistically assessing the biophysical effects of croplands.

In this study, we aim to assess the diurnal-scale biophysical impacts of croplands on Ts across tropical Africa and to investigate the underlying mechanisms. Hourly observations of Ts, along with the attribution variables α, surface downward shortwave radiation (S ↓ ), surface downward longwave radiation (L ↓ ), sensible heat flux (H), latent heat flux (LE), and ground heat flux (G), are obtained from the Meteosat Second Generation (MSG) geostationary satellites at 0.05° ×  0.05° resolution for the period 2004–202030,31. Cropland expansion in tropical Africa has predominantly occurred on former grasslands9; to ensure a consistent framework for analysing land-use transitions, this study focuses exclusively on grassland-to-cropland conversions, using a space-for-time substitution approach over a moving window32,33,34 (Methods). Croplands and grasslands are delineated using the MODIS land cover dataset35. We consider only the local biophysical effects of croplands, excluding non-local effects arising from changes in atmospheric circulation34. Given the observed spatial inconsistencies in cropland biophysical effects on Ts, we further examine their relationships with background hydroclimatic gradients. A physics-based hourly linear attribution framework is then applied to decompose cropland biophysical effects on Ts into contributions from individual biophysical factors, including α, S ↓ , L ↓ , H, LE, and G (Methods). Finally, informed by the diagnostically identified primary attributions, we employ structural equation modelling (SEM) to elucidate the biophysical processes underlying cropland impacts on diurnal Ts across tropical Africa.

Results

Observed diurnal inconsistencies in cropland biophysical impacts on Ts

Cropland biophysical impacts on Ts are quantified as the difference in Ts between cropland and surrounding grassland pixels (cropland minus grassland, ΔTs), using a space-for-time substitution approach (Methods). Strong yet contrasting cropland biophysical impacts on diurnal ΔTs across tropical Africa are observed using 17-year MSG geostationary satellite observations, as evidenced by the diurnal variations in their probability density functions (PDFs) (Fig. 1). Here, the humidity index (HI, Methods) is employed to distinguish hydroclimatic regimes, where lower values correspond to drier areas. Croplands in more arid regions (HI < 0.4) reduce ΔTs relative to surrounding grasslands, exhibiting stronger cooling at noon than at midnight. In contrast, in less arid regions (HI > 0.4), croplands increase ΔTs at noon but reduce it at midnight. Maps of cropland biophysical effects at four local solar times (00, 06, 12, and 18 LST) reveal that the strongest differences across hydroclimatic regimes occur at 12 LST, with warming and cooling effects spatially separated by a HI of 0.4 (Fig. 2). The HI threshold of 0.4 used to separate more arid and less arid regions in this study is not a prescribed hydroclimatic boundary but serves as a grouping reference to facilitate comparison between contrasting response regimes. Our main conclusions remain robust to modest shifts in this threshold (e.g., HI = 0.35 or 0.45, see Supplementary Fig. 1).

Fig. 1: Probability density functions (PDFs) of diurnal cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Fig. 1: Probability density functions (PDFs) of diurnal cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Full size image

Panels show the PDFs for (a) more arid regions (humidity index, HI < 0.4) and (b) less arid regions (HI > 0.4). The vertical lines at the top of each panel indicate the mean ΔTs across all valid moving windows. Hourly ΔTs is calculated as cropland minus grassland using the space-for-time substitution approach (Methods). Ts is derived using 17-year mean hourly Meteosat Second Generation (MSG) geostationary satellite observations.

Fig. 2: Maps of diurnal cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Fig. 2: Maps of diurnal cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Full size image

Panels show the map for (a) 00 LST, (b) 06 LST, (c) 12 LST, and (d) 18 LST, with the pink line indicating humidity index (HI) = 0.4, which separates the more arid (north) and less arid (south) hydroclimatic regimes. ΔTs at these four specific hours is calculated as cropland minus grassland using the space-for-time substitution approach (Methods). Ts is derived using 17-year mean hourly Meteosat Second Generation (MSG) geostationary satellite observations.

Similar patterns of ΔTs are observed from 20-year polar-orbiting MODIS satellite data (Supplementary Fig. 2). Daytime overpasses reveal cooling effects in more arid regions (HI < 0.4) and warming effects in less arid regions (HI > 0.4), whereas nighttime overpasses exhibit spatially consistent cooling effects. To ensure that our conclusions are not biased by spatial heterogeneity within each space-for-time substitution window, we analyse daytime MODIS observations at the four corners of each window (Supplementary Fig. 3), thereby minimizing potential biases arising from cropland and grassland samples being located in systematically different hydroclimatic backgrounds. We find that the choice of corner does not affect the overall results, with all four cases consistently showing cooling effects in more arid regions and warming effects in less arid regions during daytime (Supplementary Fig. 3). Additional seasonal analyses are conducted for the dry (December–January–February, DJF) and wet (June–July–August, JJA) seasons. The results indicate that the direction of ΔTs remains equally consistent throughout the year in both MSG geostationary satellite observations (Supplementary Figs. 4 and 5) and MODIS observations (Supplementary Fig. 6), with only a generally weaker nighttime signal during the wet season. Despite seasonal differences in HI magnitude, the similar meridional gradient (Supplementary Fig. 7) allows classification of hydroclimatic zones with distinct daytime ΔTs responses to cropland in both dry and wet seasons. Accordingly, all subsequent analyses in this study are performed on an annual basis, without accounting for seasonal variability.

Attribution of cropland biophysical impacts on diurnal Ts

To investigate how croplands biophysically influence Ts relative to surrounding grasslands, we develop a linear attribution framework (Methods). This framework provides a detailed decomposition of the contributions from individual pathways, including α, S ↓ , L ↓ , H, LE, and G. Attribution analyses are conducted on an hourly basis to capture diurnal variations. The estimated ΔTs (\(\Delta {T}_{s}^{{{\rm{est}}}}\)) from the linear attribution framework reproduces the diurnal patterns observed by satellite (\(\Delta {T}_{s}^{{{\rm{obs}}}}\)), as illustrated in the upper two panels of Fig. 3. Moreover, the diurnal variations in the PDFs of \(\Delta {T}_{s}^{{{\rm{est}}}}\) (Supplementary Fig. 8) are broadly consistent with those of \(\Delta {T}_{s}^{{{\rm{obs}}}}\) (Fig. 1) in both the more arid and less arid regions. The overall correlation coefficient between estimated and observed ΔTs is 0.54, while the slope of the regression is close to unity (0.98), indicating reasonable agreement in trends (Supplementary Fig. 9). Sign consistency analyses show that about 62% of data points share the same direction between estimates and observations, supporting the diagnostic capability of the attribution framework to capture overall spatiotemporal variations. We acknowledge remaining discrepancies due to system complexity and measurement noise. This framework is primarily intended for diagnosing the dominant biophysical pathways influencing ΔTs rather than for precise numerical prediction; therefore, validation metrics should not be overinterpreted as indicators of predictive accuracy.

Fig. 3: Attribution analyses of diurnal cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Fig. 3: Attribution analyses of diurnal cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Full size image

\(\varDelta {T}_{s}^{{\mbox{obs}}}\) is calculated as cropland minus grassland using the space-for-time substitution approach (Methods, averages of the results shown in Figs. 1 and 2). \(\varDelta {T}_{s}^{{\mbox{est}}}\) is the estimation using the linear ΔTs attribution framework (Methods). Biophysical contributions of surface albedo (α), surface downward shortwave radiation (S ↓ ), surface downward longwave radiation (L ↓ ), sensible heat flux (H), latent heat flux (LE), and ground heat flux (G) to cropland-induced ΔTs are displayed (Methods). Yellow and green bars correspond to regions with the humidity index (HI) < 0.4 and HI > 0.4, respectively. The error bars indicate the 99% confidence interval over spatial aggregation. All datasets are based on 17-year mean hourly Meteosat Second Generation (MSG) geostationary satellite observations.

In general, the results show that ΔTs across tropical Africa are primarily mediated through changes in α (Δα), changes in H (ΔH), and changes in LE (ΔLE), with minimal contributions from changes in S ↓ (ΔS ↓ ), and relatively negligible contributions from changes in L ↓ (ΔL ↓ ) and changes in G (ΔG) (Fig. 3). ΔL↓ and ΔG are sufficiently small due to similar atmospheric and soil backgrounds within the space-for-time substitution window, resulting in negligible contributions to ΔTs. The contributions of each pathway are more evident during daytime. Specifically, the combined effects of ΔH and ΔLE, representing changes in surface turbulent heat fluxes, dominate the responses, producing cropland cooling in more arid regions (HI < 0.4) and cropland warming in less arid regions (HI > 0.4). In contrast, Δα acts as a buffer, exerting opposite effects by promoting cropland warming in more arid regions (HI < 0.4) and cropland cooling in less arid regions (HI > 0.4). At nighttime, the slight cropland cooling is mainly attributable to the enhanced ΔLE.

Cropland-induced ΔH, ΔLE, and the combined changes in turbulent heat flux (ΔH + ΔLE) relative to the surrounding grasslands are presented in Supplementary Fig. 10. The maps, together with the PDFs, consistently reveal more positive changes in turbulent heat flux in more arid regions (HI < 0.4) but more negative changes in less arid regions (HI > 0.4), with these contrasting changes primarily associated with ΔLE. Further evaluation using FLUXCOM-RS data, which produces upscaled H and LE by merging FLUXNET eddy covariance measurements with remote sensing through a machine learning approach, supports the same conclusions (Supplementary Fig. 11). The buffering contributions of Δα are also clearly illustrated in its map (Supplementary Fig. 12a) and PDF (Supplementary Fig. 12d), with cropland α lower than surrounding grasslands in more arid regions (HI < 0.4) and higher in less arid regions (HI > 0.4). Meanwhile, changes in top-of-atmosphere albedo (Supplementary Fig. 12c, f) exhibit the same pattern as Δα, but with slightly smaller magnitudes due to variations in effective cloud albedo (Supplementary Fig. 12b, e). This pattern arises because changes in cloud fraction are consistent with variations in surface turbulent heat flux (Supplementary Fig. 13), resulting in cropland having higher effective cloud albedo than the surrounding grasslands in more arid regions (HI < 0.4), but lower values in less arid regions (HI > 0.4). The positive influence of surface turbulent heat flux on cloud fraction has also been reported in previous studies investigating landscape impacts on clouds34. Collectively, these spatially inconsistent variations in cloud fraction and cloud effective albedo account for the weak yet contrasting ΔS↓ responses across hydroclimatic regions (Fig. 3).

Potential mechanisms of cropland biophysical impacts on diurnal Ts

The attribution analyses in the above section indicate that ΔTs across hydroclimatic regions are closely linked to changes in three surface variables: Δα, ΔH, and ΔLE, although the direction of their effects varies. Here, we explore the mechanisms driving differences in these variables between the croplands and their surrounding grasslands, and how their combined effects shape the observed spatial patterns of ΔTs responses.

Initially, differences in leaf area index (LAI) between croplands and grasslands (ΔLAI) vary across hydroclimatic regions, with higher ΔLAI in more arid regions (HI < 0.4), but lower ΔLAI in less arid regions (HI > 0.4) (Supplementary Fig. 14). This pattern is further supported by a joint histogram of average HI within each space-for-time substitution window against ΔLAI, which shows that croplands in drier regions consistently have higher LAI than their surrounding grasslands (Supplementary Fig. 15). One possible explanation is that more intensive irrigation practices to sustain crop growth in more arid regions lead to higher ΔLAI27,29. The regression line in Supplementary Fig. 15 shows a transition of ΔLAI from positive to negative values at around HI = 0.4, which aligns with the HI threshold we use to distinguish between different hydroclimatic regions. It should be noted that the division at HI = 0.4 is used empirically for grouping purposes only and does not represent a universal or physically prescribed hydroclimatic threshold.

For the following mechanism analyses, we separate daytime (08–16 LST mean) and nighttime (20–04 LST mean) periods. A strong negative correlation is observed between ΔLE and ΔH for both daytime and nighttime, indicating an interaction between the two variables. When the total energy at the surface is limited, an increase in LE generally corresponds to a decrease in H, and vice versa (Supplementary Fig. 16). Therefore, ΔLE and ΔH are combined as changes in turbulent heat flux (ΔLE + ΔH) for further investigations. Since nighttime ΔTs is not influenced by Δα, the mechanisms at night are relatively straightforward and are discussed first. The increased ΔLAI of croplands relative to nearby grasslands in drier regions enhances the turbulent heat flux difference (Fig. 4e), thereby contributing to a cooler surface (Fig. 4f). Similar turbulent heat flux-mediated ΔTs occur during the daytime, but with larger magnitudes (Fig. 4c, d). Increased ΔLAI in drier regions is accompanied by enhanced ΔLE and reduced ΔH (Supplementary Fig. 17), and this response pattern is consistent between daytime and nighttime, although the magnitudes are larger during the daytime owing to stronger available energy. This consistent behaviour suggests that evapotranspiration-driven energy redistribution plays a dominant role in shaping turbulent heat flux differences throughout the diurnal cycle. In addition, during daytime, croplands in drier regions exhibit higher ΔLAI, which darkens the canopy and leads to reduced Δα (Fig. 4a). Reduced Δα would normally be expected to warm the surface. However, Fig. 4b shows the opposite response. This indicates that the impact of Δα on ΔTs is outweighed by concurrent changes in turbulent heat flux, resulting in a counterintuitive positive correlation between Δα and ΔTs. Indirectly, increases in Δα decrease changes in turbulent heat flux, thereby raising ΔTs, demonstrating that α-driven changes in surface energy modulate ΔTs via turbulent heat flux.

Fig. 4: Correlations among cropland-induced changes in surface properties over tropical Africa.
Fig. 4: Correlations among cropland-induced changes in surface properties over tropical Africa.
Full size image

Panels illustrate the relationships during daytime (08–16 LST mean): a changes in leaf area index (ΔLAI) versus changes in surface albedo (Δα), b Δα versus changes in land surface temperature (ΔTs), c ΔLAI versus changes in turbulent heat fluxes (ΔLE + ΔH), d ΔLE + ΔH versus ΔTs; and during nighttime (20–04 LST mean): e ΔLAI versus ΔLE + ΔH, f ΔLE + ΔH versus ΔT. The changes (Δ) are calculated as cropland minus grassland using the space-for-time substitution approach (Methods). All datasets are based on 17-year mean hourly Meteosat Second Generation (MSG) geostationary satellite observations. Scatter points are coloured according to the humidity index (HI). The linear regressions are represented by the black lines, with the texts displaying the slope of the linear fit, correlation coefficient (R), and P-value from a Student’s t test.

We further employ structural equation modelling (SEM, see Methods) to elucidate and quantify the biophysical mechanisms through which hydro-climatology affects ΔTs, separately for daytime and nighttime (Fig. 5). Our SEM incorporates four primary biophysical mediators identified in the preceding analyses: i.e., ΔLAI, Δα, ΔLE, and ΔH. All variables are standardized via the Z-score method prior to modelling. Path effects (mean ± standard error) are used to evaluate the magnitude and direction of the linkages among these variables and HI, as well as their combined influence on ΔTs. We begin with the premise that differences between croplands and grasslands are modulated by hydroclimatic conditions, whereby HI shows a negative effect on ΔLAI ( − 0.46  ±  0.006). Nighttime and daytime are then separately modelled. During nighttime, ΔLAI positively influences ΔLE (0.41  ±  0.006) but has a minimal and statistically insignificant (P-value = 0.08) impact on ΔH (0.004  ±  0.003). ΔH, in turn, is strongly determined by ΔLE ( − 0.92  ±  0.003), indicating that changes in turbulent heat flux are largely mediated by ΔLE. Ultimately, nighttime ΔTs is primarily linked to ΔLE ( − 0.98  ±  0.015). During daytime, ΔLAI negatively influences Δα (−0.70  ±  0.005) while exerting a positive effect on ΔLE (0.27  ±  0.008) and a weaker, yet still positive, effect on ΔH (0.08  ±  0.003). Both ΔLE and ΔH are also indirectly affected by ΔLAI through Δα, due to Δα-induced changes in surface radiative energy. Nevertheless, ΔH is strongly determined by ΔLE ( − 1.08  ±  0.003). These combined effects ultimately shape daytime ΔTs, which is primarily associated with ΔLE ( − 1.02  ±  0.014). Overall, the SEM results highlight both the similarity and contrast between nighttime and daytime mechanisms: in both periods, ΔTs is primarily consistent with ΔLE, yet nighttime ΔLE is largely linked to ΔLAI, whereas daytime ΔLE is directly associated with ΔLAI and indirectly modulated by Δα. This underscores the greater complexity and interplay of biophysical pathways during daytime compared to nighttime.

Fig. 5: Results of the structural equation modelling (SEM) in illustrating the pathways through which the humidity index (HI) influences cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Fig. 5: Results of the structural equation modelling (SEM) in illustrating the pathways through which the humidity index (HI) influences cropland biophysical impacts on land surface temperature (ΔTs) over tropical Africa.
Full size image

The SEMs are built via changes in leaf area index (ΔLAI), surface albedo (Δα), latent heat flux (ΔLE), and sensible heat flux (ΔH). The changes (Δ) are calculated as cropland minus grassland using the space-for-time substitution approach (Methods). All variables are standardized via the Z-score method prior to modelling. Separate SEMs are conducted for the left panel (daytime, 08–16 LST mean) and the right panel (nighttime, 20–04 LST mean). Numbers on the lines indicate the mean values of the path coefficient and 1 standard error. Statistical significance is determined using a Student’s t-test. Solid and dashed lines represent significant (P-value < 0.001) and non-significant (P-value ≥ 0.001) pathways, respectively. Goodness-of-fit index (GFI) and root mean square error of approximation (RMSEA) are reported for the SEM frameworks of both daytime and nighttime. Given the large sample size (N = 22916), the elevated RMSEA values are considered reasonable. All datasets are based on 17-year Meteosat Second Generation (MSG) geostationary satellite observations.

Conclusions

In summary, using 17 years of hourly geostationary satellite observations, this study demonstrates that cropland expansion over natural grasslands in tropical Africa has distinct effects on local surface temperatures. These effects are found to be dependent on the diurnal cycle and background hydroclimatic conditions. We quantify the diurnal ΔTs between nearby croplands and grasslands using a space-for-time substitution approach. Nighttime ΔTs generally exhibits weak cooling. In contrast, daytime ΔTs responses are much stronger but vary with hydroclimatic conditions, showing cooling in more arid areas and warming in less arid regions. These contrasting patterns reflect differences in underlying biophysical processes. A physics-based hourly linear attribution framework indicates that changes in turbulent heat fluxes are the dominant explanatory components, while daytime ΔTs are also partially moderated by Δα. These biophysical changes are closely linked to ΔLAI, which tends to be higher in more arid regions. Overall, our findings provide a process-based interpretation of how land-use changes modulate local climate across hydroclimatic gradients, highlighting the potential risk of enhanced daytime warming associated with cropland expansion in less arid regions.

Methods

Geostationary satellite observations

The main diurnal-scale analyses in this study are based on observations from the MSG geostationary satellites operated by the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT)36. Ts, α, S ↓ , L ↓ , H, LE, surface emissivity (\(\epsilon\)), and LAI are obtained from the Land Surface Analysis Satellite Application Facility (LSA SAF) products30. Following common practice, the blue-sky α is approximated as the mean of black-sky and white-sky α, given their small difference and strong correlation32. G is taken from the Satellite Application Facility on Climate Monitoring (CM SAF) LANDFLUX Ed. 1 product31, cloud fraction from the CM SAF CLAAS-3 dataset37, and effective cloud albedo from the CM SAF SARAH-3 dataset38. All datasets have a spatial resolution of 0.05° × 0.05°. Hourly values of Ts, α, S ↓ , L ↓ , H, LE, and G at full clock hours between 2004 and 2020 are extracted to construct diurnal cycles. The original data in Coordinated Universal Time (UTC) are converted to local solar time (LST) and aggregated into long-term mean (2004–2020) hourly values prior to analyses. Additionally, long-term mean (2004–2020) values of LAI, cloud fraction, and effective cloud albedo are used.

Modis observations

The level 3 monthly MODIS/Terra Ts product (MOD11C3, version 061) at a spatial resolution of 0.05° × 0.05° for the period 2001–2020 is used as an additional dataset for robustness testing39. Terra is a sun-synchronous satellite with local overpass times at approximately 10:30 a.m. and 10:30 p.m., providing complementary information for both daytime and nighttime analyses.

For delineating croplands and grasslands, we use the level 3 annual Terra-and-Aqua-combined MODIS land cover type product (MCD12C1, version 061) at a resolution of 0.05° × 0.05° from 2001 to 202035. The International Geosphere–Biosphere Programme (IGBP) classification is adopted for analyses. Following a previous study9, IGBP class 12 (croplands) and class 14 (cropland natural vegetation mosaic) are combined into a single cropland category, while IGBP class 10 (grasslands) and class 11 (permanent wetlands) are merged into a single grassland category. The land cover map of tropical Africa is shown in Supplementary Fig. 18.

FLUXCOM-RS data

For validation, additional long-term mean (2001–2020) LE and H data from FLUXCOM-RS at 0.0833° resolution are used. This dataset is generated by upscaling FLUXNET eddy covariance measurements with MODIS remote sensing through a machine learning approach40. LE and H from FLUXCOM-RS are bilinearly resampled to a 0.05° × 0.05° grid to match the spatial resolution of the other datasets.

Digital elevation model (DEM)

To minimize the influence of altitude on the estimated cropland biophysical impacts on Ts, we use the DEM from the Shuttle Radar Topography Mission (SRTM) version 4 with a spatial resolution of 90 m41. The DEM is subsequently aggregated to a 0.05° × 0.05° resolution for consistency with the other datasets.

Hydroclimatic data

We define the HI to characterize the background hydroclimatic conditions. HI is calculated as the ratio of long-term mean (2001 − 2020) annual precipitation amount to potential evapotranspiration, using monthly data from the Climate Research Unit (CRU) TS v4.08 datasets at a 0.5° × 0.5° spatial resolution42. Lower HI values indicate more arid conditions, whereas higher values correspond to more humid climates. Supplementary Fig. 7a presents the spatial distribution of HI. Finally, HI is resampled to a 0.05° × 0.05° resolution.

Space-for-time substitution

This study employs a space-for-time substitution approach to assess the biophysical impacts of croplands on Ts. This method has been widely used to evaluate the effects of land cover changes on temperature32,33, surface energy balance43,44, and cloud properties34,45. The underlying assumption is that neighbouring land patches experience similar climatic conditions, and differences in their characteristics can serve as proxies for temporal changes. The method captures only local biophysical effects.

Areas classified as unaltered land cover are defined as pixels where the dominant land cover type remains unchanged throughout 2001–2021. A moving-window approach is applied to identify comparison pairs between unaltered cropland and unaltered grassland pixels. For each moving window, we use a size of 9 × 9 pixels (0.45° × 0.45°). A window is selected if its centre pixel is either cropland or grassland and if the combined fraction of cropland and grassland exceeds half of the window area. To minimize topographic effects, the standard deviation (s.d.) of elevation within each window is calculated, and samples with an s.d. greater than 100 m are excluded following previous studies34,46, removing 57.70% of all moving windows. In total, 22916 pixels in the study area are retained for subsequent analyses. Finally, the potential biophysical effect of cropland on a given variable (ΔX) is quantified as:

$$\Delta X={X}_{{{{\rm{cropland}}}}}-{\overline{X}}_{{{{\rm{surrounding}}}} \,\, {{{\rm{grasslands}}}}}$$
(1)

or

$$\Delta X={\overline{X}}_{{{\rm{surrounding}}} \,{{{\rm{croplands}}}}}-{X}_{{{\rm{grassland}}}}$$
(2)

where Eqs. (1) and (2) correspond to cases where the central pixel of the moving window is unaltered cropland and unaltered grassland, respectively. \({X}_{{{\rm{cropland}}}}\) and \({X}_{{{\rm{grassland}}}}\) represent the multi-year mean values of the variable X over unaltered cropland and unaltered grassland pixels, respectively. \({\overline{X}}_{{{\rm{surrounding}}}\, {{{\rm{grasslands}}}}}\) and \({\overline{X}}_{{{\rm{surrounding }}}\,{{{\rm{croplands}}}}}\) denote the mean values of the surrounding \({X}_{{{{\rm{grass}}}}{{{\rm{land}}}}}\) and \({X}_{{{{\rm{crop}}}}{{{\rm{land}}}}}\,\) within a moving window when the central pixel is unaltered cropland and unaltered grassland, respectively.

While irrigation and other management activities can influence surface energy fluxes and temperature, the observation-based methods employed here do not explicitly separate biophysical effects from management interventions. However, the extensive spatial and temporal coverage helps average out localized management heterogeneity, reducing the impact of site-specific management practices on our large-scale analyses. Moreover, regions dominated by intensive irrigation often exhibit distinct climatic and vegetation patterns, which can lead to systemic changes in vegetation biophysical effects. Disentangling biophysical effects from management practices remains challenging using current remote sensing data. Nevertheless, it is important to note that cropland expansion inherently involves intensive human management activities. Thus, our results remain meaningful for understanding the overall biophysical effects associated with land conversion, even though intensive management influences are inevitably embedded in the observations. We suggest that future studies integrating detailed management datasets and ground-based measurements would be valuable to further refine our understanding of cropland expansion impacts.

Linear ΔTs attribution framework

The redistribution of energy through the surface energy balance serves as the dominant mechanism by which land use and land cover change exerts its biophysical impacts on climate47,48,49. Fundamentally, the surface energy budget framework partitions land-atmosphere interactions into key components, including α, S ↓ , L \(\downarrow\) , H, and LE, thereby representing the equilibrium of surface energy exchanges. The linear ΔTs attribution framework begins with the land surface energy balance, which can be expressed as follows:

$${S}_{\downarrow }-{S}_{\uparrow }+{L}_{\downarrow }-{L}_{\uparrow }={LE}+H+G$$
(3)

According to the Stefan–Boltzmann law, \({L}_{\uparrow }\) is proportional to the fourth power of Ts, expressed as:

$${L}_{\uparrow }=\epsilon \sigma {T}_{s}^{4}$$
(4)

where \(\sigma\) is the Stefan–Boltzmann constant (5.67 × \({10}^{-8}{{\rm{W}}}{{{\rm{m}}}}^{-2}{{{\rm{K}}}}^{-4}\)).

Additionally, α is used to replace \({S}_{\uparrow }\), and the final form of the surface energy balance equation is expressed as follows:

$$\epsilon \sigma {T}_{s}^{4}={\left(1-\alpha \right)S}_{\downarrow }+{L}_{\downarrow }-H-{LE}-G$$
(5)

The difference between croplands and grasslands, denoted as Δ, is derived by differentiating the surface energy balance equation. A first-order Taylor series expansion is performed around the mean state of each moving window, allowing ΔTs to be derived from linearized perturbations in surface energy components. The equation is shown as:

$$\Delta {T}_{s}=\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}\left[-{S}_{\downarrow ,0}\Delta \alpha +\left(1-{\alpha }_{0}\right)\Delta {S}_{\downarrow }+\Delta {L}_{\downarrow }-\Delta H-\Delta {LE}-\Delta G\right]+\varepsilon$$
(6)

where variables with subscript 0 represent the mean values within the moving window. The Taylor series linearization can introduce errors to the estimated ΔTs. Therefore, we add a residual (\(\varepsilon\)) to minimize the errors. Finally, we apply Eq. 6 to estimate ΔTs for the long-term mean diurnal cycles. The optimal \(\varepsilon\) at each hour is determined by minimizing the root mean square error (RMSE) between the estimated and observed ΔTs over all spatial grid pixels. The terms \(-\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}{S}_{\downarrow ,0}\Delta \alpha\), \(\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}\left(1-{\alpha }_{0}\right)\Delta {S}_{\downarrow }\), \(\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}\Delta {L}_{\downarrow }\), \(-\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}\Delta H\), \(-\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}\Delta {LE}\), \(-\frac{1}{4\epsilon \sigma {T}_{s,0}^{3}}\Delta G\) are, respectively, used to quantify the contributions of \(\alpha\), \({S}_{\downarrow }\), \({L}_{\downarrow }\), \(H\), \({LE}\), and \(G\) to \(\Delta {T}_{s}\).

Structural equation modelling (SEM)

While the linear attribution framework based on the land surface energy balance equation provides valuable insights into the dominant contributors to ΔTs, internal correlations among variables limit the ability to fully disentangle their individual effects. To address these interdependencies inherent in the attribution variables, we complement the decomposition analyses with SEM (implemented using the semopy package in Python)50. SEM enables simultaneous estimation of direct and indirect effects among correlated variables, offering a more nuanced understanding of the biophysical impacts of cropland biophysical impacts on ΔTs. Specifically, we hypothesize that the hydroclimatic condition, as quantified by HI, influences ΔTs through distinct pathways during daytime and nighttime. The HI is assumed to firstly affect ΔLAI. During daytime, ΔLAI influences ΔTs through its effects on Δα, ΔLE, and ΔH, whereas during nighttime, ΔLAI affects ΔTs primarily via ΔLE and ΔH. Additionally, Δα can modulate ΔLE and ΔH, and ΔLE also affects ΔH. In practice, SEM is conducted separately for daytime (08:00–16:00 LST mean) and nighttime (20:00–04:00 LST mean) across the study area. All variables are standardized using the Z-score method prior to modelling. Finally, we report the mean path coefficients along with their standard errors and associated P-values.

It should be noted that the SEM employed here are diagnostic rather than strictly causal. The SEM framework represents hypothesized statistical pathways among variables based on physical understanding, rather than direct evidence of physical causation. Nevertheless, this approach provides a useful quantitative framework to assess the relative importance of different biophysical pathways and to evaluate whether the observed patterns are consistent with established land-atmosphere interaction theories.