Biogeophysical Effects of Land-Use and Land-Cover Change Not Detectable in Warmest Month

1. Introduction

Land-use and land-cover changes (referred to simply as “land use”) affect climate biogeochemically and biogeophysically. The biogeochemical effects of land use on climate occur through interactions in the carbon cycle, mostly moving carbon from land reserves to the atmosphere. Land use also influences climate biogeophysically by modulating energy and water fluxes that vary between different vegetation types and surface characteristics.

The biogeophysical effects of deforestation on near-surface temperatures are controlled by evapotranspiration, surface roughness and albedo (Jia et al. 2019; Pongratz et al. 2021). As the sign of these effects varies across latitudinal bands and time scales, they may counteract the biogeochemical cooling effects of re/afforestation efforts. At an annual mean time scale, deforestation may cool temperatures at high latitudes (de Noblet-Ducoudré et al. 2012) and warm them in mid-to-equatorial latitudes (Duveiller et al. 2018), yet for the summer season deforestation could ubiquitously lead to warming (Lejeune et al. 2018). In the northern latitudes, this cooling is understood to be caused by the increased albedo and reduced available energy of deforested regions relative to their darker, forested origins. Importantly, in the Northern Hemisphere winter, this albedo effect is amplified. At this time, existing forest cover traps outgoing radiation through multiple reflection within the canopy, and a reduced canopy exposes larger areas of the land surface to snowfall (de Noblet-Ducoudré et al. 2012). In lower latitudes, which are less energy-limited, the temperature response of deforestation is dominated by the reduced capacity of the land surface to lose heat through turbulent fluxes. With losing vegetative cover, the lowered canopy conductance of water from soil moisture reserves decreases evapotranspiration and the latent heat flux. As well, surface roughness is reduced after deforestation, thereby reducing the dissipation of energy from the surface through turbulent sensible and latent heat fluxes (Davin and de Noblet-Ducoudre 2010; Winckler et al. 2019b). The latter of these warming mechanisms may also explain the boreal summer warming response to deforestation. In this case, the reduced turbulent heat fluxes from a decrease in surface roughness forces a temperature increase, compensating the albedo effect’s cooling (Bright et al. 2017; Duveiller et al. 2018).

These biogeophysical temperature responses to deforestation, however, are inconsistent across modeling studies using Earth system models (ESMs) (Lejeune et al. 2017, 2018). In idealized and historical model simulations, modeled temperature responses differ in magnitude (Winckler et al. 2019c; De Hertog et al. 2022) and in the latitudes where they switch from a warming to a cooling signal (Boysen et al. 2020). This is owed to uncertainties in our understanding of historical land cover change, how land use change is implemented in the models and resulting model-dependent physics for land–atmosphere interactions (Pitman et al. 2009; Meier et al. 2019, 2022). These differences are evident in model discrepancies in the simulated partitioning of net radiation into sensible and latent heat flux in response to deforestation, for example, even when they agree on the reduction of available energy and the sign of change in temperature responses to deforestation (de Noblet-Ducoudré et al. 2012; Boysen et al. 2020).

Comparisons between modeled and observed biogeophysical temperature responses to deforestation are challenging. Often, models may simulate cooling responses to deforestation while observations show warming (Winckler et al. 2019a). This difference stems from how model simulations at the grid scale and observations taken at single points in the landscape are characterizing the local and nonlocal responses of deforestation at different spatial scales. Local effects refer to atmospheric changes at the site of a land disturbance. Yet, the atmospheric conditions over a given plot of land may also be affected by the atmospheric feedbacks of land use occurring in other regions, i.e., the nonlocal effects. Some have used the “space-for-time” concept to describe the trade-offs in the assessment of this signal between models and observations. This is because observational studies, by construction, sometimes only include the local signal. For example, in situ and satellite observations may compare a local plot of deforested land to neighboring forests to assess this signal at a single point in time. In this context, the effects of internal climate variability, including their influence on the nonlocal atmospheric interactions of land use, are constant across neighboring sites and are therefore assumed to cancel (Lee et al. 2011). These assumptions no longer hold in many studies assessing the temperature response of deforestation from ESM simulations with grid scales greater than 100 km. Comparisons of historical simulations without land use against those with time-evolving land use scenarios ingest both the local effects and the nonlocal feedbacks of deforestation at the grid cell level. Further observation-model discrepancies in the case of satellite assessments stem from an inappropriate comparison of variables. Satellites measure surface temperatures, while most model studies assess near-surface or 2-m air temperatures (Bright et al. 2017; Duveiller et al. 2018). Indeed, comparisons of modeled surface and near-surface temperature responses to deforestation reveal a disagreement in their magnitudes and sometimes in their sign of change (Winckler et al. 2019c; Breil et al. 2020; Pongratz et al. 2021). Finally, Chen and Dirmeyer (2020) reconcile these differences between observations and a single model. In running offline simulations to avoid feedbacks and by evaluating them to mimic the space-for-time conditions of observational studies, they replicate the observed warming in response to deforestation. It is thus likely that the observation–model discrepancies in the temperature response to land use are not only due to model imperfections, but also because the nonlocal effects of land use are less accounted for in the observations.

The local and nonlocal effects of deforestation have only been separated in the model world. This is achieved by reconstructing historical deforestation signals and by performing idealized simulations. Reconstruction methods for assessing the signals of historical deforestation have been designed in cases where counterfactual historical simulations—those including all historical forcings but which hold land use constant—were not available to derive the land use signal. With modeled temperature and tree cover information, these applications use either regressions (Winckler et al. 2019c; Lejeune et al. 2018; Thiery et al. 2020) or mean differences between high and low tree cover change grid cells (Kumar et al. 2013; Lejeune et al. 2017) in a rolling window of grid cells to parse the temperature changes owing to varying tree cover. Importantly, this also filters some nonlocal effects under the assumption that they are shared in each window. This is evidenced by the reduction of uncertainty in the reconstructed temperature signal of deforestation when compared to factorial derivations of the signal (i.e., a lowered influence of internal variability, which modulates nonlocal effects) (Lejeune et al. 2018). Idealized “checkerboard”-type deforestation analyses apply a methodology to extract local against nonlocal effects of deforestation by using composite fields of forested and deforested grid cells (Winckler et al. 2017; De Hertog et al. 2022). These studies show that models simulate summertime local warming and nonlocal cooling. Similarly, in idealized deforestation studies that apply global deforestation rates not expected in any reasonable future land use scenario, it is thought that the nonlocal cooling masks local warming effects (Boysen et al. 2020). While evapotranspiration efficiency and surface roughness changes drive the local response (Winckler et al. 2017), albedo changes are thought to be the causal factor behind this nonlocal cooling. Albedo changes affect the temperature and humidity of the whole troposphere, allowing for advection of these perturbed conditions to other regions (Davin and de Noblet-Ducoudre 2010).

In addition to improving our process understanding of the biogeophysical effects of land use, resolving the significance of its contribution to historical climate change is of key importance. Yet this question of detection and attribution of the effects of land use on climate has received relatively little attention (Christidis et al. 2013). Detection and attribution involve assessing the significance of a signal in some aspect of the climate, and then evaluating the roles of different climate forcings that might affect it (Hegerl et al. 2006). This has been undertaken extensively, for example, to evaluate the role of anthropogenic greenhouse gas emissions in affecting target variables like temperature and precipitation relative to natural forcings, such as volcanoes and solar activity (Zwiers et al. 2011; Wan et al. 2015). Detection is defined as first demonstrating that a target variable has changed in a statistically significant way. This significance is defined relative to the distribution of possible changes in the target variable as forced by internal climate variability. Therefore, a signal is detected if it can be confidently determined that its probability of occurring under internal climate variability is sufficiently small. Upon detecting a signal, questions remain over which climate forcings were responsible for its significant shift. Attribution is then the process of determining the suspect forcings responsible for this shift against other plausible explanations (Hegerl et al. 2007).

One challenge in the context of detecting and attributing the effects of land use on climate is the potential nonlocal cooling response to deforestation. For locations and time scales relevant to the nonlocal cooling response, this conflicts with the spatially uniform carbon sequestration benefits of re/afforestation mitigation efforts. Existing evidence implies that this cooling may be forced by nonlocal processes linked to internal climate variability. Therefore, exploring the significance of the historical temperature response to land use in a way that takes into account internal variability is helpful. Methodologically, detection and attribution involves ensemble sampling and optimization steps that reduce the influence of internal climate variability on the historical land use signal.

Recent advances in land use scenarios (Hurtt et al. 2020) and the availability of historical simulations with and without land use change (Lawrence et al. 2016), which can dissect the biogeophysical effects of land use in phase 6 of the Coupled Model Intercomparison Project (CMIP6), merit a revisiting of the detection and attribution of the overall land use signal. Previously, a single-model study found an attributable cooling of historical warmest day temperatures to land use (Christidis et al. 2013). The researchers used an earlier generation of UKESM that included a dynamic vegetation model that ingested agricultural disturbance data from early land use harmonization datasets (Hurtt et al. 2011). However, considering the discrepancies in modeled temperature responses to land use, a more complete estimate of the land use signal’s detectability may only be achieved in a multimodel framework. Here we revisit the detection and attribution of land use effects on climate by focusing on 2-m warmest month temperatures (the locally highest annual value of monthly means of daily maximum temperatures) for four ESMs simulated under the updated protocol of CMIP6 (Eyring et al. 2016).

2. Data and methods

a. Models and observations

For detection and attribution, we require observations and ESM simulations of historical and preindustrial climates (Ribes et al. 2013). Model simulations come from the following CMIP6 concentration-driven experiments: all historical forcings (“historical,” hereafter HIST), all historical forcings without land use (“hist-noLu,” hereafter HISTNL), and a preindustrial control representing the effects of internal variability (“piControl,” hereafter PIC) (Eyring et al. 2016). Here, concentration-driven means that models are forced by scenario-prescribed CO₂ concentrations instead of calculating atmospheric CO₂ explicitly within an ESM’s carbon cycle. The use of simulations from concentration-driven experiments isolates the biogeophysical effects of land use (Lawrence et al. 2016). For observations, we use Berkeley Earth monthly land temperature fields (BE; Rohde and Hausfather 2020) and Climatic Research Unit gridded Time Series version 4.04 (CRU; Harris et al. 2020), which span the historical period and provide gridded temperature fields for monthly means of daily maximum temperatures at higher resolutions than the model fields (Table 1).

Table 1

Observational and model-based dataset characteristics.

View Table

ESMs include the Canadian Earth System model version 5 (CanESM5; Swart et al. 2019), the second generation Earth System model from CNRM-CERFACS (CNRM-ESM2-1; Séférian et al. 2019), the Earth system model from the Institut-Pierre-Simon Laplace (IPSL-CM6A-LR; Boucher et al. 2020) and the U.K. Earth System model (UKESM1-0-LL; Sellar et al. 2020). These models were selected on the basis of data availability on the public Earth System Grid Federation repository. Each model must have at least three realizations available in both HIST and HISTNL experiments for monthly means of daily maximum 2-m temperature. Hereafter, this variable is referred to as tasmax, and the response of tasmax to the forcings included in HIST and HISTNL experiments is denoted t_HIST and t_HISTNL, respectively. Multiple realizations are required for computing individual model means to reduce the effects of internal climate variability. Per model, the maximum number of realizations of t_HIST and t_HISTNL were included (Table 1). For details on the guidance of land use application in participating CMIP6 models, see Lawrence et al. (2016), Hurtt et al. (2020). Many studies focus on warm rather than annual mean or cold temperature responses to land use. Typically, surface property changes drive a stronger radiative response in the surface energy balance with more solar insolation (Bright et al. 2017; Duveiller et al. 2018; Hirsch et al. 2015; Lejeune et al. 2018). We therefore opt for tasmax over other temperature indices, as this expresses a stronger signal and enriches existing daily warm temperature attribution to land use (Christidis et al. 2013). Using the monthly frequency is necessary due to data availability constraints. Note that tasmax should respond to land use the strongest at the daily frequency, yet few ESMs had sufficient outputs at daily frequency for both HIST and HISTNL to consider within this analysis (see section 4).

b. Preprocessing

We compute the mean representation of individual ESM experiments before space-time aggregation. For each ESM, we compute the mean t_HIST and t_HISTNL time series across available realizations (Table 1). Reconstructing an experiment for the effects of land use (LU) on tasmax (t_LU) is then derived by subtracting the mean t_HIST by the mean t_HISTNL. For capturing internal climate variability, we gather as many nonoverlapping PIC realizations as possible that match the length of our study period. These PIC realizations are taken from all CMIP6 models with simulations of our target variable (t_PIC). From 39 models and 25 893 years of PIC data, this yielded 498 realizations of t_PIC. This developed from originally confining the sampling of t_PIC from only the ESMs in Table 1. However, this limited the quality of internal variability estimates in the regression due to small sample sizes. We therefore opt for a compromise; assuming that all CMIP6 models adequately simulate preindustrial internal variability and that the sample across many models suitably represents the internal variability in the simulations of our four ESMs, we improve the quality of internal variability estimates in our analysis.

We select the most recent 50 years of the CMIP6 historical period (1965–2014) from all observed and modeled input data. This invokes a trade-off in our analysis; although 1915–64 would more strongly express the effects of land use relative to other historical forcings, selecting the more recent period improves the reliability of the observations and land use trajectories in the Land-Use Harmonization project (LUH2) (Hurtt et al. 2020). Time series for 1965–2014 are then resampled to 5-yr nonoverlapping block means.

In anticipation of the different ESM responses to land use and in hope of preserving the local features of these responses, we avoid regridding all models and observations to a common resolution. Instead, we regrid observations to match the spatial resolution of the individual models. This avoids otherwise (i) promoting redundancy in ESM t_LU patterns by interpolating coarser model fields to the finer observational resolutions, or (ii) dampening the ESM signal expression of t_LU through resampling all model and observational fields to a coarse resolution (Lejeune et al. 2017). This means that for individual ESM detection and attribution results, the observations have been regridded to match the resolution of the ESM’s simulations. For the multimodel mean results and for pooling t_PIC data across many models, space–time aggregation is done before computing ensemble averages, thereby neglecting regridding.

Spatial preprocessing of all model and observational data is undertaken for coarsening inputs to detection and attribution to best express the signal to noise ratios of land use forced temperature changes. All input data points to the regression are confined to land regions, excluding Greenland, the Sahara, and Antarctica. The data points are area-weighted spatial averages that are bounded depending on the spatial scale of the detection and attribution result. A first set of weights takes into account different grid cell areas by scaling their contributions to the regional average by the cosine of their latitude. At the continental scale—where separate detection results are achieved for North America, South America, Europe, Asia, Africa, and Australia—inputs are aggregated according to the Sixth IPCC Assessment Report reference land regions (AR6; Iturbide et al. 2020). This contains the local effects of land use to regions of consistent climatological characteristics. Note that we include some regions that have undergone no land use on the basis that they may experience the nonlocal effects of land use. For assessing global detectability, to lower the dimensionality of the regression, spatial averages are computed per continent, excluding Antarctica. At both the continental- and global-scale regressions, we therefore use input vectors or fingerprints that capture space–time characteristics at a finer spatial resolution than the detection result. This allows for the estimated inverse variance–covariance matrix to optimize input vectors to the regression by emphasizing regions of low variability. As these subregions have different land areas, we additionally weight their data points (the area-weighted spatial averages) based on the areas of the subcontinental AR6 or continental regional domains (Fig. 1). We confine our study to land regions to avoid the influence of less certain nonlocal effects via atmospheric feedbacks over the oceans (Davin and de Noblet-Ducoudre 2010).

View Full Size

Regions used for spatial aggregation and their relative weights. (a) Continental detection analyses, yielding results separately for North America, South America, Europe, Asia, Africa, and Australia, use aggregated data for land-based grid cells based on the IPCC AR6 reference regions within each continent (continental domains are outlined additionally in black), wherein each region is given a relative weight. This splits Africa north of the Sahara, as the Mediterranean AR6 region belonging to Europe encompasses land regions of both Europe and Africa. (b) For global detection analyses, yielding a single result, spatial aggregation is contained according to ESRI definitions of continental bounds, while also excluding Greenland, the Sahara and Antarctica. For the global scale, the northern tip of Africa is not considered a part of Europe but is included in the area-weighted average for data over Africa.

Citation: Journal of Climate 36, 6; 10.1175/JCLI-D-22-0391.1

• Download Figure
• Download figure as PowerPoint slide

The space-time dimensionality reduction yields observed and model-based input vectors of size n = 10 (time steps) × 6 (continents) for performing the regression analysis on a global basis. For the continental-scale detection results, n varies according to the continent being assessed. Here, n is therefore of size 10 × AR6_c. Before the regression, we temporally center each region’s time series by removing its mean across time.

3. Results

a. Detection and attribution

1) The detectability of $t_{\text{LU}}$ in the two-way regression

The response of tasmax to land use forcings, t_LU, is inconsistently detectable in observations (BE) across ESMs at both continental and global scales, but its detectability is internally consistent across scales for specific ESMs (Fig. 2). For CanESM5, only South America contains a detectable t_LU signal (Fig. 2b), and at the global scale t_LU is undetected and indistinguishable from internal climate variability (Fig. 2c). Likewise, for UKESM1-0-LL, Africa is the only continent with a detectable signal (Fig. 2k), yet, globally, t_LU is undetectable (Fig. 2l). Across ESMs, Africa is the most consistent continent with a detectable t_LU signal, as all but CanESM5 simulate a detectable signal there. CNRM-ESM2-1 and IPSL-CM6A-LR have the most detectable signal of t_LU. At the global scale, their scaling factors are consistent with unity (Figs. 2f,i), which is apparent in their continental scale results in that they both simulate a detectable t_LU signal in at least four of six continents (Fig. 2e). The inconsistency in the signal of t_LU across ESMs is confirmed in the multimodel mean case (Figs. 2m–o), where no continents contain a significant expression of t_LU and its signal is undetectable globally. All other forcings (t_HISTNL) are nearly unanimously detected in all continents, across scales and ESMs. When repeating the analysis with CRU as an alternative set of observations, some of the continents and ESMs with detectable signals are different (Fig. A1), where, for example, IPSL-CM6A-LR loses its detectability of t_LU. However, the same outcomes around (i) inconsistent signals across ESMs, (ii) internally consistent signals propogating through continental and global scales for individual ESMs and (iii) null results for the multimodel mean case all remain true.

View Full Size

Continental and global detection and attribution results for two-way analyses with observations from Berkeley Earth. The continental scale detectability of (a),(d),(g),(j),(m) t_HISTNL (purple) and (b),(e),(h),(k),(n) t_LU (green) fingerprints. Light colors represent scaling factor results implying detection (β_f ± 90% CI > 0), dark colors represent scaling factor results consistent with unity (${\mathit{\beta }}_f\pm 90\%\text{CI}\ni 1$), and gray implies an undetected signal. (c),(f),(i),(l),(o) The global detectability of t_HISTNL (purple; x axes) and t_LU fingerprints (green; y axes) through their scaling factors and confidence intervals. In these panels, dark, hashed lines are drawn to mark unity and light gray lines mark 0. Multimodel mean results are labeled as “mmm”.

Citation: Journal of Climate 36, 6; 10.1175/JCLI-D-22-0391.1

• Download Figure
• Download figure as PowerPoint slide

2) Does $t_{\text{LU}}$ affect the detectability of the historical climate change signal?

The response patterns of tasmax to the forcings in HIST and HISTNL, t_HIST and t_HISTNL, represent alternative fingerprints of the historical climate with and without the effects of land use, and their consistency with observations (BE) is sometimes affected by the presence of t_LU in t_HIST (Fig. 3). All ESMs produce t_HIST and t_HISTNL signals that are detectable in the observations at the global scale, and Australia is the only continent where, in some instances, these signals are not detected. For the ESMs that produce a detectable signal of t_LU in the two-way regression, the presence of t_LU in t_HIST improves the detectability of t_HIST relative to t_HISTNL for some continents in CNRM-ESM2-1 (Figs. 3d,e), IPSL-CM6A-LR (Figs. 3g,h), and UKESM1-0-LL (Figs. 3j,k). However, for the multimodel mean results at both continental and global scales, there are no differences in the detectability of t_HIST and t_HISTNL. This feature is consistent for analyses with both observational datasets (Figs. 3m–o, A2 m–o).

View Full Size

As in Fig. 2, but for one-way analysis results reported separately for t_HIST (browns) and t_HISTNL (blue-greens).

Citation: Journal of Climate 36, 6; 10.1175/JCLI-D-22-0391.1

• Download Figure
• Download figure as PowerPoint slide

b. S/N of treeFrac in t_LU

The S/N of treeFrac in t_LU is generally weak (Fig. 4). The S/N for t_LU exists mostly around the center of S_PIC/N distributions (the control), and, with the exception of the tropics in CanESM5, they only reach so far as the 66th percentile of the control distribution. CNRM-ESM2-1 and IPSL-CM6A-LR have the best pattern similarity between treeFrac and t_LU, as the S/N for t_LU always exceeds the 66% percentile of the control distribution (Figs. 4b,c), which is consistent with these ESMs simulating t_LU with the best detectability using ROF for detection and attribution. In line with the two-way regression results, CanESM5, IPSL-CM6A-LR and UKESM1-0-LL have the best pattern matching between treeFrac and t_LU in the tropics to temperate-south latitude band (Figs. 4a,d), which contains the continents for which they report detection using both BE and CRU as observations (with exception to detecting UKESM1-0-LL in CRU). In all ESMs except for CNRM-ESM2-1, the tropical to temperate-south latitude band shows the likeliest signal of treeFrac in t_LU, implying that the local effects of land use are most strongly expressed in these regions. For the tropical to temperate-south latitude band, it is contradictory that CNRM-ESM2-1 produces the lowest S/N of treeFrac in t_LU while t_LU for this ESM is also consistently detectable in southern continents across observational datasets using ROF (see section 4). Since CNRM-ESM2-1 has the smallest sample of t_PIC realizations from which to compute N, it is possible that the sampled N is considerably larger than its true value, suppressing S/N for this ESM. In the other latitude bands, the lesser pattern similarity between treeFrac and t_LU may imply that noise or nonlocal effects of land use are dominant. Overall, these results emphasize the weak signal case of t_LU because the treeFrac S/N is consistent with preindustrial control runs, namely that t_LU probably contains spurious trends from internal climate variability.

View Full Size

(a)–(d) Signal-to-noise ratios (S/N) of (e)–(h) EOF fingerprints of treeFrac in t_LU vs t_PIC realizations for each model. Histograms showing the distribution of S_PIC/N across boreal, temperate-north, and tropical to temperate-south latitudinal bands are plotted behind S/N for t_LU. The S_PIC/N ratios derived from t_PIC are sampled for each latitude band across all four models in the detection and attribution analysis for computing the following quantiles for signal detectability: 66th, 90th, and 99th.

Citation: Journal of Climate 36, 6; 10.1175/JCLI-D-22-0391.1

• Download Figure
• Download figure as PowerPoint slide

4. Discussion and conclusions

We find (i) a null detection and attribution result for t_LU—the response of the local warmest monthly mean of daily maximum temperature to land-use and land-cover changes (land use)—using regularized optimal fingerprinting (ROF) in a two-way regression and (ii) no evidence that land use improves the detection of the historical forced response pattern. Even though results vary across ESMs, for the multimodel mean analysis, t_LU is consistently undetected and indistinguishable from internal climate variability in both two-way and one-way regressions and across observational datasets. CNRM-ESM2-1 is the exceptional model in the ensemble. If taken alone as a means of communicating the significance of the t_LU signal, its detectability across scales and observational datasets, reinforced to a minor degree by its pattern matching in the EOF-fingerprinting exercise, might be convincing evidence of detection. Alongside its detection results, opposite to CanESM5, the t_LU response pattern of CNRM-ESM2-1 is the most aligned with treeFrac changes north of the tropical to temperate-south latitude band. Altogether, this may imply a greater expression of local effects of land use in this ESM, thus agreeing with earlier assessments of the scale dependence of the land use signal (Chen and Dirmeyer 2020). Notably, excluding IPSL-CM6A-LR, ESMs in our ensemble simulate above average mean cooling responses to an idealized deforestation experiment (Table 1; Boysen et al. 2020). However, as discussed below, the t_LU response pattern for CNRM-ESM2-1, like the rest of the ensemble, should be viewed with caution given the difficulties in assessing why its parameterization scheme should be trusted over others for computing surface water and energy fluxes due to land use.

We also caution against extending comparisons between the results of our ROF detection and attribution and the EOF-fingerprinting analyses, as these methods address different questions. ROF for detection and attribution asks whether or not t_LU is significantly detectable in observed changes in tasmax. EOF-fingerprinting, however, seeks to test how patterns of change in in treeFrac agree with t_LU. Although we can infer from a strong S/N of treeFrac in t_LU that this could imply better signal detectability of t_LU in ROF, this only comments on the potential signal expression in t_LU. It says nothing about t_HISTNL, the observations, and how they are expressed after space–time truncation and optimization with t_PIC variance–covariance matrices. As well, this inference assumes that land use will have a stronger signal relative to all of historical climate change if it is mostly made up of local rather than nonlocal effects. This assumption may not be true in ROF if regions with strong nonlocal effects are not aligned with those of high variance in t_PIC simulations from the entire CMIP6 ensemble.

The null outcome of this analysis is acceptable for a number of reasons. For some ESMs, the inclusion of t_LU in t_HIST in the one-way regression analysis improves the consistency of modeled response patterns with observations. The opposite does not occur, wherein the inclusion of t_LU reduces the likeness between observations and models (Fig. 3, Fig. A2). Yet, the cancellation of t_LU signals in the multimodel mean underscores model uncertainties that are too high for this set of forcings relative to internal climate variability to allow for formal detection and attribution, yielding null results across scales and observational datasets. This questions the assumption that these models are all valid in representing land use. The weak S/N assessment of treeFrac in t_LU and the fact that the regions with the strongest trends in t_LU do not consistently align with regions of high land use (Figs. 4, B1) supports that factorial estimates of t_LU may be dominated by nonlocal processes and/or internal climate variability. The t_LU response pattern in our study period is a weak signal in comparison to the earlier historical period due to the stronger effects of land use relative to other forcings in the early twentieth century. This is apparent in more consistent correlations and slightly greater S/N when repeating these analyses for 1915–64 (Figs. B2, B3). Finally, we tested multiple other approaches to our detection and attribution analysis, which all provided similar results. These tested alternatives included (i) separately assessing summer and winter seasonal tasmax; (ii) making selections of regions of high land use with respect to LUH2-prescribed tree cover change before aggregating with AR6-weighted means for continental analyses; (iii) limiting the inclusion of AR6 regions to those of considerable land use; (iv) testing other temporal aggregation methods, such as either computing 10-yr means or eliminating the temporal dimension of the analysis altogether by representing each region by its difference between the means of 1965–89 and 1990–2014; and (v) restricting our t_PIC samples to those simulated by the four ESMs involved in this study.

Relative to the analysis of daily temperature extremes in Christidis et al. (2013), it is understandable that we find a weaker signal case using the monthly time scale. As they discuss regarding their attribution of warm day extremes relative to the insignificant effect of land use on cold and night extremes, ample solar insolation is required to drive the expression of the radiative effects of land use that control temperature changes. This includes albedo controlling the available energy and the partitioning of this energy into turbulent heat fluxes. Therefore, on the one hand, in dulling the temporal scale of this analysis from daily to monthly resolution, the peak assessment of these radiative effects on near-surface temperatures is lost, effectively weakening our version of the land use signal. On the other hand, the signal to noise ratio of t_LU in our study is benefited by analyzing the monthly resolution, which is less influenced by internal variability than daily temperatures. However, comparisons between this study and Christidis et al. (2013) can never be perfect. We isolated spatial aggregations to land regions and avoided ocean domains, while they used a T4 truncation that included ocean domains. Repeating our analysis to include the oceans, however, invites potentially greater inconsistencies in ESM-specific advection processes and ocean–atmospheric coupling. Although we choose a likewise 50-yr-long study period, ours begins 15 years later, which means we are including different land use trends in some regions. This study is therefore not a correction to their detection and attribution results, but rather an addition to the outlook of the detectability of land use on warm temperature indices.

Our detection and attribution analysis has a number of limitations. Our observational datasets, CRU and BE, likely do not fully ingest historical tasmax responses to land use. By convention, weather stations do not measure areas before and after land use transitions such as deforestation. Moreover, in CRU, tasmax is estimated using daily mean temperature and the diurnal temperature range (Thiery et al. 2020). The same occurs on the model side, wherein daily tasmax is computed at the model time step. As these time steps do not necessarily align with the timing of the peak temperature, the daily maximum temperature might be underestimated by the ESMs. However, as ESM time steps typically are between 30 min and 1 h, we assume that the potential effect is limited. Model data availability constraints also limit the quality of this analysis. There are too few ESMs involved in this study to robustly assess t_LU detectability across the full CMIP6 ensemble. As our four-ESM ensemble displays weak and spatially varying trends in t_LU and different applications of land use (Figs. B1, B2), encompassing all CMIP6 model outcomes could provide a more reliable understanding of the significance of land use as a historical forcing. A limited availability of ensemble members for the HISTNL experiment leaves our estimate of t_LU contaminated by internal variability (Table 1). Similarly, restricting preindustrial control runs to only those simulated by the ESMs involved in this study would cover too few years for adequately sampling the space-time variance possible via internal climate variability. Therefore, we increase this sample size as a compromise for potentially including internal variability estimates that are ill-fitted to our four ESMs. In this way, the assumed internal variability is representative of a CMIP6 multimodel mean variability, although that internal variability varies largely among CMIP6 models in terms of amplitude and characteristics (Ribes et al. 2021). Accounting for these between-model differences may inflate the variance of internal climate variability for some regions, thereby reducing the likelihood of detection and making our detection test more conservative. Finally, we apply an earlier subset of the ROF regression toolbox by using an ordinary least squares (OLS) fitting of model predictands to observed predictors. In the detection and attribution context, this ignores uncertainty in the model fingerprints by minimizing the vertical distances of the residuals in the regression, which is best imagined in the one-way case (Fig. 2; Allen and Stott 2003). Methodological advances from Allen and Stott (2003) came by the use of total least squares (TLS) fitting, which assumes the fingerprints to be a noised version of their true pattern and involves fitting the regression by minimizing distances perpendicular to the best fit line. In our case, the OLS regression acts as a conservative approach to searching for a weak signal. Our inability to reject the null hypothesis that t_LU is indistinguishable from noise in the multimodel mean case using OLS implies that further work with the more robust TLS fitting is not yet necessary.

Our EOF-fingerprinting analysis demonstrates the presence of internal climate variability and/or nonlocal effects of land use in t_LU and benefits from its basis in the treeFrac fingerprint. This approach lends a pattern assessment framework from earlier studies assessing the presence of modeled fingerprints in observations for the same variable (Marvel et al. 2019; Santer et al. 2018). We deviate from this by evaluating the likeness between the tree cover fingerprint, a prominent variable in the land use forcing (Lejeune et al. 2017), and t_LU. In the context of the effects of land use—originating local to the site of land perturbation yet affecting other regions due to advection and nonlocal feedbacks (Winckler et al. 2017)—the pseudoPC series from projecting t_LU onto the treeFrac fingerprint indicates the degree to which patterns of change in t_LU match those of tree cover. Therefore, this also indicates if our tasmax metric for 1965–2014 is dominated by local or nonlocal processes. Since the S/N of treeFrac in t_LU is mostly consistent with the control distribution, it cannot be claimed that t_LU is composed mostly of local responses to land use. Instead, t_LU is likely influenced by internal climate variability or the nonlocal effects of land use. Importantly, t_LU also contains forced responses to other types of land use, such as agricultural expansion and urbanization. Therefore, features in t_LU that might appear to be nonlocal feedbacks to forest cover change will sometimes be local responses to other land perturbations. However, tree cover changes are known to produce a stronger climatic signal than crop cover expansion (Lejeune et al. 2017), and our results did not change much when testing this approach for cropFrac. The presentation style of these results is typical in detection and attribution studies. A metric derived from different scenarios, be it trends in or correlations between response patterns and a historical mean series, is compared against a control distribution derived from preindustrial control simulations (Qian and Zhang 2015; Padrón et al. 2020; Grant et al. 2021). However, applying the same approaches in this study would not achieve much given the weak and inconsistent signal of t_LU. In our approach, despite that there is also implicit uncertainty behind the land cover change data that the ESMs apply, on an individual ESM level, the patterns of the treeFrac fingerprints act as concrete basis against which to assess t_LU.

There are many sources of uncertainty in t_LU which control its inconsistency across ESMs. ESMs respond differently at the local scale to land use, owing to their individual systems for representing vegetation (Sellar et al. 2020; Séférian et al. 2019; Swart et al. 2019; Boucher et al. 2020) and for computing surface flux exchanges for water and energy cycles (de Noblet-Ducoudré et al. 2012; Lejeune et al. 2020). For example, it is interesting to note that the models without dynamic vegetation schemes have the best expressed signals of land use. These local responses to land use are then propagated nonlocally by model-specific advection and atmospheric feedback processes, which are inconsistent across ESMs in this study and in earlier research (Pitman et al. 2009; Winckler et al. 2019c; Hirsch et al. 2015). One such process includes cloud cover simulation, a known limitation to modern ESMs which has a demonstrated cooling bias from overestimating cloud shortwave radiative effects (Miao et al. 2021). Future work could distinguish the drivers of model uncertainties between their surface flux responses or atmospheric feedbacks by comparing the coupled simulations applied here to the historical Land Use Model Intercomparison Project land-only simulations with and without land use (Lawrence et al. 2016). Meanwhile, it is beyond the scope of this study to assess what drives these differences in temperature outcomes in detail. This would require factorial experiments with incremental changes to understand at a process level what may drive differences between ESMs, a challenging task due to cascading interactions. Finally, the basis for these t_LU response patterns, the land use patterns themselves, are differently applied in the models owing to model-dependent land parameterization and tiling schemes. This includes how the number of PFTs vary between the land surface schemes and whether or not they include land management. The background information from LUH2, which prescribes the locations and timing of historical land use, also has demonstrated uncertainties and limitations (Hartung et al. 2021; Prestele et al. 2017).

To conclude, our detection and attribution analysis found an insignificant influence of biogeophysical land use on the daily maximum temperatures averaged over the warmest month, t_LU. This temperature metric is the best available climate variable across CMIP6 ESMs for expressing the influence of this set of forcings on warm temperatures given the data requirements of a detection and attribution exercise. A null result was found confidently by assessing t_LU across different observational datasets, ESMs, spatial scales and analytical approaches to detection and attribution. This result, when viewed through the lens of the extensive research showing the difficulty of representing land use impacts in ESMs, is not necessarily surprising. This study clarifies that, given existing model imperfections, it is not yet possible to find a significant fingerprint of the effects of land use on warm temperatures via traditional detection and attribution methods that rely on accurately modeled response patterns.

Data availability statement.

Data from models is available under the CMIP6 archive on the Earth System Grid Federation search engine here: https://esgf-data.dkrz.de/search/esgf-dkrz/. Observations from Berkeley Earth can be found here: http://berkeleyearth.org/data/. Observations from the Climatic Research Unit can be found here: https://catalogue.ceda.ac.uk/uuid/c26a65020a5e4b80b20018f148556681. Code for this analysis is here: https://github.com/VUB-HYDR/2022_Grant_etal_JOC.