A New Approach to Soil Initialization for Studying Subseasonal Land‐Atmosphere Interactions

Numerical experiments on sensitivity to land surface initializations are frequently conducted to investigate the predictability and uncertainties of hydrometeorological extremes. However, the conventional approaches to soil initialization often assume synchronized extremes over the target region, creating initial conditions that violate the intrinsic spatial pattern of hydrometeorological variability. Here we propose a “Slope” approach to accommodate unsynchronized anomalies, which creates initial conditions of a variable (soil temperature or soil moisture) across a large domain based on the slopes of linear regression between the variable averaged over a small target region and at each grid point in the surrounding regions. Within the target region, the “Slope” approach produces spatial patterns and temporal evolutions of hydrometeorological responses similar to the conventional approach, but generates stronger signals probably due to the nonlocal impact (excluded from the conventional approach). In the surrounding regions, the hydrometeorological responses in the “Slope” approach are consistent with the spatiotemporal variability of the model climate. Slope‐based experiments targeting different adjacent regions produce similar results, suggesting that one ensemble of experiments targeting one region may be sufficient to represent the responses from multiple ensembles each targeting a different region and thus providing the basis for increasing the computational efficiency of some land‐atmosphere interaction studies. While South America is used to demonstrate the concept in this study, the new approach offers the most advantages in regions with spatially unsynchronized or even anti‐phased hydrometeorological extremes.

Wetter soil tends to increase evapotranspiration (ET), cool the land surface, and favor local rainfall through modifying atmospheric convection and moisture convergence (Seneviratne et al., 2010).From an energy balance perspective, wetter soil enhances the surface net radiation through the direct impact of lower surface albedo and the indirect impact of lower Bowen ratio, leading to more heat flux into the atmosphere, a higher density of moist static energy within the atmospheric boundary layer, and potentially more precipitation (Eltahir, 1998).Through these mechanisms, wetter soil can enhance rainfall thus becoming wetter, leading to a positive feedback (e.g., Kim & Wang, 2007a).While an overwhelming body of literature suggests a positive feedback, soil moisture can also facilitate a negative feedback to precipitation (Cook et al., 2006;Guillod et al., 2015;Leutwyler et al., 2021;Taylor et al., 2012;Yang et al., 2018).For example, Taylor et al. (2012) found that afternoon precipitation falls preferentially over drier soil compared to the surrounding area, and the negative feedback is attributed to the enhanced moist convection driven by increased sensible heat flux over drier soil in the daytime, which highlighted the importance of spatial heterogeneity in soil moisture.The discrepancies in the direction of soil moisture-precipitation feedback among past studies have to do with not only the dependency on geolocations but also the sparse observational records for soil moisture and the analytical techniques used to identify feedback (e.g., spatial vs. temporal analyses) (Guillod et al., 2015;Tuttle & Salvucci, 2016).Moreover, the land-atmosphere coupling strength differs substantially among different models (Koster et al., 2004(Koster et al., , 2006;;Orlowsky & Seneviratne, 2010;Taylor et al., 2011).There is a clear need to further improve the representation and our interpretation of land-atmosphere feedback in coupled models (Taylor et al., 2012;Tuttle & Salvucci, 2016).
Due to the carryover effect, a soil moisture anomaly can be "remembered" in the land-atmosphere system and can thereby influence subsequent atmospheric processes at the subseasonal timescale (Dirmeyer & Halder, 2017;Koster & Suarez, 2001;Koster et al., 2020;Liu et al., 2014).Sensitivity experiments on land surface initialization using numerical models are frequently conducted to study land-atmosphere feedback and to identify the sources of predictability and uncertainty related to hydrometeorological extremes (e.g., Kim & Wang, 2007a, 2007b;Koné et al., 2022;Koster et al., 2010;Koukoula et al., 2021;Saini et al., 2016;Seo et al., 2019), and various approaches to soil moisture initialization have been used.Earlier studies applied homogeneous soil moisture (usually a fixed value) over the target region to initialize the coupled land-atmosphere models (e.g., Yeh et al., 1984), which may lead to misrepresentation of heat flux partitioning and precipitation (Georgescu et al., 2003;Santanello et al., 2019).To account for the spatial heterogeneity of soil texture and soil moisture, some studies applied a spatially uniform level of initial soil saturation or uniform anomalies (Asharaf et al., 2012;Grimm et al., 2007;Kim & Wang, 2007a, 2007b, 2012;Kutty et al., 2018;Meng & Quiring, 2010;Pal & Eltahir, 2001;Zhong et al., 2018Zhong et al., , 2019)).For example, Pal and Eltahir (2001) derived multiple levels of soil moisture saturation (10%, 25%, 50%, 75%, and 90%) over the entire U. S. Midwest from observations, and used those to initialize the land surface scheme of a regional climate model in simulating the 1988 drought and the 1993 flood.Yet another approach to account for the spatial heterogeneity is to set the initial soil moisture of a target region to a certain percentile individually derived for each grid cell based on long-term statistics from model simulations or observational data set (e.g., Liu et al., 2014;Meng & Quiring, 2010;Sun & Wang, 2011, 2012).For example, Sun and Wang (2012) prescribed initial soil moisture at the 10th, 50th, and 90th percentiles of a control simulation.Although both the uniform saturation level and the uniform percentile of soil moisture are more realistic than homogeneous soil moisture in the study domain, a common and fundamental issue with these previous approaches is the implicit assumption that the variabilities or extremes of soil conditions are synchronized across a large study domain.In the real world, dipole pattern or anti-phased variations is common; soil moisture or temperature may be higher than normal in one portion of the study domain and lower than normal elsewhere (Littlefield et al., 2020;Marengo et al., 2013).Uniform or synchronized anomalies may occur over a small region only.However, imposing anomalies to a small region creates a sharp contrast across its boundaries and similarly violates the intrinsic spatial pattern of the coupled land-atmosphere system.
To achieve realistic spatial heterogeneity over a large domain, here we propose a "Slope" approach to soil temperature and moisture initializations for coupled land-atmosphere modeling at the subseasonal timescale.Relative to conventional approaches, the new approach facilitates a more realistic simulation of the land-atmosphere system by preserving the slopes of linear regression between soil temperature or moisture in the target region and its surroundings.Section 2 describes relevant components of the model used, the soil initialization approaches, and the experimental design.Comparisons between the conventional and the "Slope" approaches are presented in 10.1029/2023MS003822 3 of 16 Section 3. Section 4 shows the efficiency of the "Slope" approach, followed by a summary and discussion in Section 5.

Model Description
In this study, we used the Community Earth System Model version 2.1 (CESM2.1)(Danabasoglu et al., 2020) to conduct global simulations with sea surface temperature and sea ice fraction prescribed according to observations from the Hadley Centre and the National Oceanic and Atmospheric Administration (Hurrell et al., 2008).The atmospheric component for the coupled land-atmosphere model is the Community Atmosphere Model version 6 (Bogenschutz et al., 2018), and the land component is the Community Land Model version 5 (CLM5) with satellite phenology (D.M. Lawrence et al., 2016Lawrence et al., , 2019)).Relative to the previous version of the land model, the soil hydrology scheme in CLM5 is largely improved, including better simulation of ET and total water storage due to a new mechanism-based parameterization for soil evaporation resistance and a dry land surface layer constraining the rate of water vapor diffusion (Swenson & Lawrence, 2014).CLM5 also includes a total of 25 unevenly spaced vertical ground layers, and the first 5 (0.4 m thickness) up to 20 (8.5 m thickness) layers are soil and hydrologically active (Brunke et al., 2016;D. M. Lawrence et al., 2019).The remaining ground layers in each column are bedrock.Although CLM5 has the capacity to simulate vegetation dynamics, in this study we prescribed the vegetation structure, distribution, and phenology according to observations (Lawrence & Chase, 2007).The leaf area index (LAI) of each plant functional type (PFT) varies from day to day, but the LAI seasonal cycle and PFT coverages do not vary from year to year, and remain at the level of the mean over 2001-2015.To exclude the uncertainties originating from other potential sources, atmospheric CO 2 concentration was set at 367.0 ppm and trace gas concentrations were fixed at their year 2000 levels.

Soil Initialization Approaches
The spatially uniform percentile approach to soil initialization is chosen as an example of the conventional approaches, following the method of Liu et al. (2014) and Sun andWang (2011, 2012).This conventional approach only imposes soil anomalies to the target region and soil conditions elsewhere remain the same as control.Over the target region, soil conditions at each percentile and each soil layer can be derived from a long time series.Due to the absence of a sufficiently long observational record in tropical South America, the soil temperature and moisture statistics are derived from a long model simulation in this study.Specifically, take the soil temperature in September as an example, the initial soil temperature for September 1 is obtained from daily soil temperature data during August 16-September 15 in each of 34 years during 1985-2018.That is, a total of 1054 (31 × 34) days of soil temperature data for each soil layer at each grid point are sorted separately to derive the 1st and 99th percentiles over the target region, and these percentile threshold values are used to initialize soil temperature in the target region.Since no soil anomalies are imposed in the surrounding regions, this conven tional approach creates a sharp contrast across the boundary of the target region.
In the real world and in climate models as well, the soil temperature and moisture variations in the surrounding regions tend to follow a certain spatial pattern relative to the target region.Figure 1 presents the linear regression slopes (and the corresponding correlation coefficients) of the soil temperature and moisture between the spatial average over the target regions (Southern Amazon and Nordeste) and all grid points over tropical South America based on monthly output from the long simulation.While the correlation between the target region and all grid points are strong and significant over most of the study domain (except for the Andes where the correlation is not significant), the slope of the regression shows a large spatial heterogeneity across tropical South America.For example, soil temperature anomalies in Nordeste and the south of 20°S have opposite signs in September (Figure 1a2).Although the majority of tropical South America tends to have anomalies of the same sign, the magnitude varies substantially across the region.The approximately linear relationship in Figure 2 further demonstrates that when soil temperature or moisture in Southern Amazon is above normal, the soil conditions in Nordeste also tend to be above normal too.Even though some data points are in the second and fourth quadrants for both soil temperature and moisture, none of them represents an extreme case.
Given the strong correlation over most of the domain (Figure 1) and the linearity of the relationship (Figure 2), we propose to derive the soil anomalies in regions surrounding the target region based on the linear regression 10.1029/2023MS003822 4 of 16 slope using a three-step procedure, explained in the following using the 99th percentile of Tsoi for September 1 as an example.
1. Soil temperature initialization in the target region: The 99th percentile soil temperature for each soil layer at each grid point (  99_grid ) of the target region (e.g., Nordeste) is derived directly from the daily output of a 34-year long simulation during August 16-September 15 (1054 daily data points in total).2. Estimation of the linear regression slopes: The monthly spatial average of the top 10-cm soil temperature over the target region is calculated based on output from the 34-year long simulation (in total, 34 independent monthly data).The linear regression slopes S (i,j) between this monthly spatial average and the monthly soil temperature at every grid point elsewhere in South America are then derived.Note that due to its slow variation, daily soil variables (daily soil moisture in particular) feature strong autocorrelation, which may result in spurious regression slopes.This nonphysical effect of autocorrelation can be eliminated if the data is aggregated temporally (e.g., weekly or monthly).In this study, we derived the regression coefficients based on monthly averaged data; sensitivity analysis results indicated that the regression coefficients based on weekly data are largely similar.3. Soil temperature initializations in the surrounding regions: The September 1 temperature anomalies for each grid point are first derived within the target region  ∆target_grid = 99_grid − CTL_grid ; these anomalies are then The derived anomalies are then used to construct initial soil temperatures in the surrounding regions that conform to the intrinsic spatial heterogeneity as compared to those in the target region.The derivation of initial soil moisture anomalies follows a similar procedure.We refer to this newly proposed approach as the "Slope" approach.

Experimental Design
As mentioned in the Section 2.2, we first performed a continuous long global simulation (as control) at f09 horizontal resolution (approximately 0.9° × 1.25°) over the period 1980-2018, with the first five years for model spinning up.This simulation provides the data needed for estimating the soil temperature and moisture statistics (Figures 1 and 2), serves as a reference for comparison, and provides the initial conditions for multiple sensitivity experiments.
To assess the performance of the "Slope" approach, multiple global simulations that differ in initial soil conditions were conducted, including 1-month simulations initialized with different soil conditions targeting humid Southern Amazon (75°W-50°W, 17°S-5°S) and semiarid Nordeste (50°W-34°W, 17°S-0°).Our experiments did not target the Northern Amazon (75°W-50°W, 5°S-5°N), the wettest hydrological regime over tropical South America where the magnitude of soil moisture variability is low and soil saturation level is high even during drought extremes (Jiang et al., 2022).Previous studies have indicated that the climate system in tropical South America is the most sensitive to land surface conditions during the pre-monsoon season (September-October-November) (Fu & Li, 2004;Jiang et al., 2021;Li & Fu, 2004).As the tropical South American monsoon generally starts during a time window from late September to October (Fu et al., 2013), short simulations initialized on September 1 and November 1 are used to represent different hydrometeorological conditions in the study domain.
The details of the experimental design are listed and described in Table 1.Numerical experiments (Hot, Cool, Wet, and Dry) were initialized with soil conditions constructed through the "Slope" approach each year during 2006-2015.We took an ensemble approach and each ensemble has 10 members for both September and November.The simulated differences between Hot and Cool experiments (Tsoi tests) can quantify the hydrometeorological responses to soil temperature change; differences between Dry and Wet experiments (Soilliq tests) represent the impacts of soil moisture differences.As the baseline for comparison, Cool and Wet references were conducted, each initialized with the soil conditions generated from the conventional approach, using 2015 as a case study.Due to the abnormally high temperature and low soil water content in 2015 over most of the tropical South America (Erfanian et al., 2017;Jiang et al., 2022), the control simulation in fact represents an extreme hot/ dry case.For both the "Slope" and conventional approaches, differences in simulated results between control and Cool quantify the response to Tsoi in 2015; similarly, differences in simulated results between control and Wet quantify the response to Soilliq.In addition to assessing the initialization methodology, this experimental design also allows us to examine the local versus nonlocal effects of initial soil anomalies, which has implications on the potential role of land surface feedback in shaping the spatial variability or contrast of hydrometeorological conditions within the study domain.
In order to distinguish the signal from background noise, the signal-to-noise ratio (Giorgi & Bi, 2009;Hawkins & Sutton, 2012;Kent et al., 2015) is calculated for each hydrometeorological response.The ensemble mean of hydrometeorological response is a measure of the forced signal, and the standard deviation of the hydrometeorological response among the 10 ensemble members is a measure of the random noise.The hydrometeorological response is deemed to be statistically significant if the signal-to-noise ratio is larger than 1.The significance of the response can also be quantified using the student t-test.The results based on the signal-to-noise ratio and the 10% t-test are largely similar, although the latter identified a slightly smaller area with significant responses.

Constructed Initial Soil Conditions Through the "Slope" Approach
Taking the "Slope" approach, we constructed the soil initial conditions for experiments targeting Southern Amazon ("SouA experiment" hereafter) and Nordeste ("Nord experiment" hereafter) (Figure 3).The spatial patterns of soil temperature over three subregions show slight differences between the SouA and Nord experiments (Figures 3a1-3b4), and differences in spatial averages are generally smaller than 2°C (Table 2).In particular, the spatial pattern and magnitude of initial soil temperature over Northern Amazon (outside the target region) constructed for the SouA experiment closely resemble those for the Nord experiment.In the Dry experiment, the initial soil moisture magnitude and spatial pattern are remarkably similar between the SouA and Nord experiments (Figures 3c1,3c2,3d1,and 3d2).
The high similarities between the SouA and Nord experiments over three subregions (Table 2) may be partly due to the small variability in the extremely dry soil condition at the 1st percentile.More importantly, it confirms that the "Slope" approach can construct initial soil conditions with realistic spatial heterogeneity over the whole study domain, which has little to do with the choice of a target region.However, the differences tend to be larger in the Wet experiment.The differences in initial soil moisture are modest in September in the eastern part of Northern Amazon, the southern part of Southern Amazon, and the northern part of Nordeste (Figure 3c3 vs. Figure 3c4), and are larger in November after the South American monsoon onset (Figure 3d3 vs. Figure 3d4).This disparity reveals a limitation of the Slope approach: based on climatologically coherent spatial correlation, the "Slope" approach is better at capturing large-scale systematic extremes (e.g., drought or persistent wet conditions), and is not expected to capture soil moisture anomalies resulting from localized randomly occurring extreme precipitation.

Hydrometeorological Responses
Here we use 2015 as a case study to illustrate the differences in simulated hydrometeorological responses between the two initialization approaches (Figures 4 and 5). Figure 4  anomalies in the Tsoi tests and soil water anomalies in the Soilliq tests from the "Slope" approach against those from the conventional approach over the targeted region (Southern Amazon) in the first month following the initial soil anomalies.In the Tsoi tests, the soil temperature anomalies can only last for 2 weeks (Figure 4a), and the temporal variabilities of soil temperature anomalies are quantitatively similar between the two approaches.In contrast, the soil moisture anomalies in the Soilliq tests tend to persist longer than 1 month (Figure 4b).The simulated temporal variabilities of soil moisture anomalies in the Soilliq tests are similar between the two initialization approaches.The signal in the "Slope" approach is slightly stronger than the conventional approach, which may be due to the impact of nonlocal soil moisture anomalies.However, experiments using the conventional approach pertained to the year 2015 and did not include sufficient ensemble members to support a significance test.
Since the soil temperature anomalies last for no more than 2 weeks, we focus our spatially distributed analysis of simulated hydrometeorological responses on the first 10-day averages in both the Tsoi and Soilliq tests.From a technical perspective, the simulated soil temperature and moisture conditions may quickly deviate from their initial values depending on the strength and direction of land-atmosphere interactions, although due to soil memory, the impact of initialization is expected to be non-negligible especially in the first couple of days.Over the target region, the simulated soil temperature in the Tsoi tests and soil moisture in the Soilliq tests are qualitatively similar between the two approaches in both September and November (Figures 5a1-5b2 and 5g1-5h2).Quantitatively, a stronger signal of soil temperature is found in the "Slope" approach.In the Tsoi tests, the positive temperature anomalies over the target region in the "Slope" approach (1.01°C; Figure 5a1) are slightly stronger than those in the conventional approach (0.89°C; Figure 5a2) in September.However, that is not the case for November, as negative temperature anomalies are simulated in the southeast of the target region in the "Slope" approach (Figure 5b1 vs. Figure 5b2).In the Soilliq tests, the positive soil temperature responses in September and November averaged to 1.12°C and 0.49°C for the "Slope" approach, and 0.93°C and 0.25°C for the conventional approach respectively (Figures 5c1-5d2).The modest quantitative differences in the simulated soil temperature in the target region may be attributed to the nonlocal impact of soil temperature anomalies imposed in the surrounding regions in the "Slope" approach.In climate models or the real world, the hydrometeorological conditions in the target region are influenced by both the local and nonlocal initial anomalies, which can be captured by the "Slope" approach but cannot be captured by the conventional approach.In addition, the "Slope" approach could simulate the positive signal over both the target region and its surroundings (Figures 5a1, 5b1, 5c1, and 5d1), while the signal is absent outside the target region in the conventional approach (Figures 5a2,5b2,5c2,and 5d2).The long simulation suggests that soil temperature over Nordeste is expected to be above normal too when soil temperature in Southern Amazon is above normal (Figure 2a), which confirms that the "Slope" approach produces more realistic (spatially more coherent) hydrometeorological conditions in the surrounding regions.
The soil moisture responses in the Tsoi tests are mixed and very small in magnitude (Figures 5e1-5f2), indicating that soil moisture in this region is not directly driven by soil temperature but is likely dominated by the rainfall signal.In the model or real world, higher soil temperature and lower soil moisture generally co-occur (e.g., Grimm et al., 2007;Jiang et al., 2022), which can be a result of drought leading to excessive heat (as opposed to the other way around) in this region.In the Soilliq tests, negative anomalies are imposed to initial soil moisture directly, and a clear negative signal remains over most of the target region (Figures 5g1-5h2).Averaged over the target region, the negative signals in the "Slope" approach (−3.90 mm in September and −6.38 mm in November; Figures 5g1  and 5h1) are stronger than those in the conventional approach (−3.40 mm in September and −5.15 mm in November; Figures 5g2 and 5h2).In addition, while the "Slope" approach produces a negative signal over both the target region and its surroundings, the conventional approach shows a negative signal in the target region only and little signal elsewhere, an indication that nonlocal impact of soil moisture anomalies from the target region is rather small or even negligible.In the long control simulation, soil moisture over Nordeste decreases when soil moisture in Southern Amazon decreases (Figure 2b), a spatial pattern that can only be captured by the "Slope" approach due to the lack of nonlocal impact in the conventional approach.
Initial soil temperature anomalies cause negligible ET responses in September due to water limitations; in November when water is no longer limiting, warm anomalies lead to a slight increase of ET (Figures 5i1-5j2).Dry anomalies in September cause a local decrease of ET that is strong within the target region in both approaches and weak in the surrounding regions even in the "Slope" approach; in November, the ET response to initial soil moisture anomalies becomes weaker (than in September) within the target region due to the relaxed water limitation, and stronger in the surrounding region with the "Slope" approach due to its moderate water availability (Figures 5k1-5l2).The spatial patterns of evaporative responses are similar between the two soil initialization approaches within the target region.The ET responses in the surrounding regions simulated with the "Slope" approach result mostly from the local impact of initial soil moisture anomalies, as the response simulated with the conventional approach is negligible outside the target region, creating a sharp contrast across the borders of the target region (Figures 5k2 and 5l2).Despite the differences in ET response, the simulated precipitation responses are similar between the "Slope" approach and the conventional approach (results not shown).The simulated precipitation responses are generally weak and mixed over tropical South America in both the Tsoi and Soilliq tests and in both September and November.

Efficiency of the "Slope" Approach
In addition to the more realistic spatial pattern of initial conditions and the subsequent hydrometeorological responses, another benefit of the "Slope" approach is that it enables a more efficient experimental design in some studies.When accounting for the spatial correlation of hydrometeorological anomalies, separate experiments targeting different subregions that behave similarly may produce similar results.If so, one group of experiments targeting one subregion of a domain may be sufficient to represent the results from multiple experiments targeting different subregions.Here we examine the similarity in soil memory length and hydrometeorological responses among experiments targeting different subregions of South America using the "Slope" approach.Subregions that behave similarly can be identified according to the coefficients of regression and correlation (Figure 1).Here Southern Amazon and Nordeste are chosen as examples.
Figure 6 compares the temporal variations of soil temperature in the Tsoi tests and soil moisture in the Soilliq tests in three subregions from the SouA experiment (targeting South Amazon) against those from the Nord experiment (targeting Nordeste).Consistent with the findings in Figure 4, the soil temperature anomalies only last for 2 weeks, while the soil moisture anomalies can persist longer than 1 month.In the Tsoi tests, temporal variations of soil temperature anomalies in all three subregions show a clear similarity between the SouA and Nord experiments in both September and November (Figures 6a1,6a2,6b1,6b2,6c1,and 6c2).In the Soilliq tests, despite a high degree of similarity in the temporal variation, there are substantial differences in the magnitude of soil moisture anomalies between SouA and Nord experiments (Figures 6a3,6a4,6b3,6b4,6c3,and 6c4).Understandably, the soil moisture anomalies are generally larger within the target region than its surroundings in most cases.For example, in Southern Amazon, the soil moisture anomalies in the SouA experiment are more than twice as large as in the Nord experiment in both September and November (Figures 6b3 and 6b4).
All subsequent hydrometeorological variables (soil temperature, soil moisture, ET, and precipitation) in the SouA and Nord experiments show similar spatial patterns (Figures 7 and 8).In the Tsoi tests, soil temperature features a significant positive signal over tropical South America for all experiments, as the positive temperature anomalies are imposed at initialization (Figures 7a1-7b2).However, a clear contrast between the SouA and Nord experiments in September can be found in the region south of 20°S (Figure 7a1 vs. Figure 7a2).In the Soilliq tests, both the magnitude and spatial patterns of the simulated soil temperature responses are similar between SouA and Nord experiments (Figures 7c1,7c2,7d1,and 7d2).This is not without exception.For example, soil temperature responses in southeast Brazil show a clear contrast between the two September experiments (Figure 7c1 vs. Figure 7c2).
Soil moisture signal in the Tsoi tests is constrained by the local hydrological regime and precipitation responses (Figures 7e1-7f2).Soil moisture responses in the SouA experiment closely resemble those in the Nord experiment.
In September, a significant negative signal is found in Nordeste and the eastern part of Southern Amazon, while ET increases in response to positive soil temperature anomalies (Tsoi tests) and decreases in response to negative soil moisture anomalies (Soilliq tests) in both the SouA and Nord experiments, although they do not pass the significance test over most grid cells of tropical South America (Figures 8a1-8d2).In the Tsoi tests, some slight differences in spatial patterns of ET responses can be found between the SouA and Nord experiments.For example, a decrease of ET is found in the region south of 20°S in the Nord experiment in September, in contrast to an increase in the SouA experiment (Figure 8a1 vs. Figure 8a2).In November, due to the monsoon onset, ET responses are stronger than in the dry season.The discrepancy between the SouA and Nord experiments mainly concerns the western part of Southern Amazon (Figures 8b1 and b2).In the Soilliq tests, the spatial patterns of evaporative responses are generally similar between the two experiments, although some magnitude differences do exist (Figures 8c1-8d2).
The precipitation responses simulated by the SouA and Nord experiments are similar in both the magnitude and spatial pattern, and this statement holds for both the September and November experiments, and for both the Tsoi and Soilliq tests (Figures 8e1-8h2).In the dry September prior to monsoon onset, strong signals of precipitation are found only in the humid region north of 10°S for all experiments, including Northern Amazon and the northern part of Southern Amazon.In November after monsoon onset, rainfall substantially increases; the SouA and Nord experiments show a similar positive signal over tropical South America in the Tsoi tests, and a similar positive signal in the eastern Southern Amazon and western Nordeste in the Soilliq tests.
The high degree of similarity between the results from the two "Slope" experiments targeting different subregions confirms that one "Slope" experiment would be sufficient in capturing the hydrometeorological responses across most of the domain.We acknowledge the caveat that, quantitatively, one cannot interpret the Nordeste results in the SouA experiment as the response to the local 99th percentile of anomalies in Nordeste.Qualitatively, in studying the physical mechanisms of land-atmosphere interactions, a single "Slope" experiment would be sufficient.Arguably, one conventional experiment applying 99th percentile anomalies over a combined Nordeste-Southern ensemble means.All simulations were initialized with the "Slope" approach, targeting Southern Amazon (red) and Nordeste (blue), respectively.Shaded regions in all panels represent the 95% confidence intervals.
Amazon target zone could yield similar results.However, one would have to first conduct the regression analysis we proposed for the Slope approach (as shown in Figure 1) to identify areas that can be potentially combinable (as these are not prior knowledge); otherwise, a combined experiment assuming synchronized extremes would most likely violate the intrinsic spatial pattern of the coupled land-atmosphere system.

Summary and Discussion
Conventional approaches to soil initialization in land-atmosphere interaction studies often assume synchronized variability or extremes within a target region, which may not be valid depending on the domain of study.The initial conditions applied in those studies may violate the intrinsic spatial pattern in the model or in reality and fail to account for unsynchronized and sometimes anti-phased variations.To address this deficiency, here we propose a new "Slope" approach to soil temperature and moisture initialization for coupled land-atmosphere modeling at the subseasonal timescale.The "Slope" approach accounts for the linear relationship between the soil temperature or moisture averaged over the target region and their values at all grid points across the domain, generating initial conditions with realistic spatial heterogeneity.Qualitatively and within the target region, the new approach produces spatiotemporal variability of the hydrometeorological responses that are similar to the conventional approach; quantitatively the new approach generates a stronger signal over the target region likely due to the nonlocal impact of initial anomalies imposed in the surroundings.In the surrounding regions, the hydrometeorological responses in the "Slope" approach are consistent with the spatial patterns in the climate model, partly because the initial soil conditions are constructed for consistency with what is imposed in the target region.
With our "Slope" approach, experiments targeting different regions that behave similarly (e.g., Southern Amazon vs. Nordeste) share similar initial soil conditions, similar temporal variations of soil anomalies, and similar hydrometeorological responses.Such similarities suggest that one ensemble of experiments targeting one subregion may be sufficient to represent the results from multiple ensembles each targeting a different subregion.
The "Slope" approach therefore has the potential to increase the efficiency of experimental design and make ensemble-based land-atmosphere interaction studies computationally more feasible.
Deriving the regression coefficient is an important step of the Slope approach and can be a source of uncertainty.
The high degree of autocorrelation in daily soil moisture data could largely reduce the number of independent pairings for soil statistics and result in spurious regression slopes.To ensure data independence, the regression coefficients were derived based on monthly data in this study.For this reason, the "Slope" approach can only capture major systematic events and may fail to represent some extreme cases especially extremely wet conditions

Figure 1 .
Figure1.Regression coefficient (a1-b4) and Pearson correlation coefficient (c1-d4) of the top 10 cm soil temperature and moisture between the spatial average over the target region (Southern Amazon (first and third column) and Nordeste (second and fourth column)) and all grid points in tropical South America, based on monthly data derived from the 34-year long control simulation.The correlation maps only show the areas where the values pass the 5% significance test.

Figure 2 .
Figure 2. Scatter plot between monthly averaged soil conditions at top 10 cm in Southern Amazon and Nordeste.Each data point represents the spatial average of the standardized anomaly in soil temperature (a; in °C) and moisture (b; in mm) over two regions from the 34-year long control simulation.

Figure 3 .
Figure3.Constructed initial soil temperature (a1-b4; in °C) and moisture (c1-d4; in mm) in the top three soil layers (12 cm) for experiments targeting Southern Amazon (first and third column) and Nordeste (second and fourth column) using the "Slope" approach.The black rectangles in each panel mark the target region for each simulation.

Figure 4 .
Figure 4. Simulated top 10 cm soil temperature differences (a; in °C) from the Tsoi tests (Control minus Cool) and soil moisture (b; in mm) from the Soilliq tests (Control minus Wet), averaged over the target region (Southern Amazon) in September (dash line) and November (solid line) of 2015; Red and blue colors indicate results from experiments with the "Slope" and conventional approaches, respectively.

Figure 6 .
Figure 6.Top 10 cm soil temperature responses (first and second column; in °C) from the Tsoi tests (Hot minus Cool) and soil moisture responses (third and fourth column; in mm) from the Soilliq tests (Dry minus Wet) averaged over each of three subregions in September and November, based on 10-member (2006-2015)ensemble means.All simulations were initialized with the "Slope" approach, targeting Southern Amazon (red) and Nordeste (blue), respectively.Shaded regions in all panels represent the 95% confidence intervals.

Figure 7 .
Figure 7. Ensemble means of top 10 cm soil temperature responses (in °C) and soil moisture responses (in mm) from the Tsoi tests (Hot minus Cool) and Soilliq tests (Dry minus Wet), respectively.All simulations were initialized with the "Slope" approach for each September and November of 2006-2015 and targeting Southern Amazon and Nordeste, respectively.Stippling indicates areas where the signal-to-noise ratio is larger than 1.

Figure 8 .
Figure 8. Ensemble means of ET and precipitation responses (in mm/day) in the Tsoi tests (Hot minus Cool) and Soilliq tests (Dry minus Wet), respectively.All simulations were initialized with the "Slope" approach for each September and November of 2006-2015 and targeting Southern Amazon and Nordeste.Stippling indicates areas where the signal-to-noise ratio is larger than 1.

Table 1
List of 1-Month Short Simulations compares the spatially averaged soil temperature

Table 2
Constructed Initial Soil Temperature and Moisture in the Top Three Soil Layers (12 cm), Averaged Over the Subregions of Tropical South America