Deforestation reshapes land-surface energy-flux partitioning

Land-use and land-cover change significantly modify local land-surface characteristics and water/energy exchanges, which can lead to atmospheric circulation and regional climate changes. In particular, deforestation accounts for a large portion of global land-use changes, which transforms forests into other land cover types, such as croplands and grazing lands. Many previous efforts have focused on observing and modeling land–atmosphere–water/energy fluxes to investigate land–atmosphere coupling induced by deforestation. However, interpreting land–atmosphere–water/energy-flux responses to deforestation is often complicated by the concurrent impacts from shifts in land-surface properties versus background atmospheric forcings. In this study, we used 29 paired FLUXNET sites, to improve understanding of how deforested land surfaces drive changes in surface-energy-flux partitioning. Each paired sites included an intact forested and non-forested site that had similar background climate. We employed transfer entropy, a method based on information theory, to diagnose directional controls between coupling variables, and identify nonlinear cause–effect relationships. Transfer entropy is a powerful tool to detective causal relationships in nonlinear and asynchronous systems. The paired eddy covariance flux measurements showed consistent and strong information flows from vegetation activity (gross primary productivity (GPP)) and physical climate (e.g. shortwave radiation, air temperature) to evaporative fraction (EF) over both non-forested and forested land surfaces. More importantly, the information transfers from radiation, precipitation, and GPP to EF were significantly reduced at non-forested sites, compared to forested sites. We then applied these observationally constrained metrics as benchmarks to evaluate the Energy Exascale Earth System Model (E3SM) land model (ELM). ELM predicted vegetation controls on EF relatively well, but underpredicted climate factors on EF, indicating model deficiencies in describing the relationships between atmospheric state and surface fluxes. Moreover, changes in controls on surface energy flux partitioning due to deforestation were not detected in the model. We highlight the need for benchmarking model simulated surface-energy fluxes and the corresponding causal relationships against those of observations, to improve our understanding of model predictability on how deforestation reshapes land surface energy fluxes.


Introduction
The land surface and atmosphere are closely coupled through energy, water, and carbon cycles (Santanello et al 2013, Gentine et al 2019. For example, the land surface dissipates energy and water into the atmosphere through evapotranspiration and latent heat fluxes. Plants open stomata to receive atmospheric CO 2 and produce carbohydrates, which simultaneously lose water though transpiration (Lei et al 2010, Suyker andVerma 2010). The dissipated latent heat and evapotranspiration fluxes modulate atmospheric temperature and water vapor pressure, which in return can affect vegetation and soil processes (Allison andTreseder 2011, Lombardozzi et al 2015).
The strengths of such two-way interactions between land and atmosphere are strongly dependent on landsurface characteristics (e.g. soil temperature, wetness) (Dirmeyer 2011, Ford et al 2014, Feldman et al 2019, and atmospheric conditions (e.g. vapor pressure deficit (VPD), radiation, and air temperature) (Zhang et al 2014, Zhou et al 2014, Kukal and Irmak 2016. Moreover, land-atmosphere interactions are largely mediated by local vegetation (e.g. grass or tree) and exhibit complex energy, water, and carbon coupling among the soil, the vegetation, and the atmosphere (Puma et al 2013, Williams andTorn 2015).
To better understand the land-atmosphere coupling that includes the whole soil-vegetationatmosphere continuum, two segments of this coupling are commonly distinguished: (a) the land segment, i.e. the linkage between land states (e.g. soil moisture) and surface energy flux patterns (e.g. latent heat flux) and (b) the atmosphere segment, i.e. the relationship between surface-energy flux patterns and atmospheric states (e.g. precipitation) (Guo et al 2006, Bowling et al 2010. In both land and atmospheric segments, the evaporative fraction (EF = LE/(LE + H), where LE and H are latent and sensible heat fluxes) is considered a central variable of interest (Koster et al 2009, Ford et al 2014, Feldman et al 2019, which defines how the land surface partitions net radiation into latent versus sensible heat fluxes. Surface energy partitioning and land-atmosphere coupling are subject to land-surface characteristics (Koster et al 2004, Carleton et al 2008, Koster et al 2010 and can be affected by land-use and landcover change (LULCC) (Cooley et al 2005, Hirsch et al 2014, Lorenz and Pitman 2014. As one of the major LULCC activities, deforestation breaks or at least weakens the existing linkage within the land-vegetation-atmosphere continuum, by removing woody plants that could potentially transpire a large amount of water into the atmosphere (Claussen et al 2001, Davin and de Noblet-ducoudré 2010, Myoung et al 2012. Consequently, the land surface tends to dissipate energy through the sensible heat pathway (Gash and Nobre 1997) rather than the latent heat pathway in deforested versus forested land (Evaristo and McDonnell 2019). However, observed EF does not necessarily decline after deforestation, owing to the concurrent changes of albedo and surface energy balance (Luyssaert et al 2014). The separation of the compounding changes (vegetation cover and background atmospheric forcings) (Winckler et al 2017) is one of the major challenges in understanding how deforestation affects both land-surface energy partitioning and land-atmosphere coupling.
The patterns and controllers of EF have been widely investigated in the scientific literature (Brimelow et al 2011, Findell et al 2011. However, most of these studies have focused on the impacts related to soil states (i.e. soil moisture) (Ford et al 2014, Feldman et al 2019, while fewer studies have considered the role of vegetation cover and other factors (Myoung et al 2012, Williams et al 2016. This knowledge gap necessitates a comprehensive examination of how vegetation cover and other environmental states could potentially impact surfaceenergy partitioning before and after deforestation. A second challenge is to extend the analysis of the soil moisture-EF relationship (Dirmeyer 2011) to include relevant variables that could potentially drive surface-energy partitioning.
Linear correlative relationships have been widely used to analyze coupling between land states (i.e. soil moisture) and surface-energy partitioning (Basara andCrawford 2002, Ford et al 2014), even though nonlinearity in these interactions are acknowledged (Brubaker 1995, Zeng et al 2002. Therefore, linear-correlation-based representation of landatmosphere coupling potentially leads to an incomplete understanding or misleading conclusions. In this study, we analyze these nonlinearities and develop relevant metrics to quantify nonlinear impacts on surface-energy partitioning from deforestation. We diagnose the strength of landatmosphere coupling using a recently synthesized FLUXNET dataset (Chen et al 2018) with paired forest-versus-non-forested land surfaces that share common climate forcing. Thus, differences in landatmosphere coupling strength between paired forested and non-forested sites can be attributed to deforestation-induced land-cover change. We apply a novel and nonlinear metric (transfer entropy) based on information theory (Liu et al 2019) to diagnose how land-cover change (deforestation) affects EF and land-atmosphere coupling.

Paired FLUXNET dataset
In this study, 29 pairs of FLUXNET sites from the FLUXNET2015 Tier 1 and AmeriFlux datasets (Baldocchi et al 2001) were used. The spatial locations of the paired sites are shown in figure 1, and the general information of the sites are listed in table S1 (available online at stacks.iop.org/ERL/16/024014/mmedia). Each pair contains a forested (e.g. broadleaf or needleleaf forests) site and a nearby non-forest (e.g. grassland, cropland, or open shrub) site. The median linear distance between the paired sites is 21.6 km, and the median elevation difference is 20.0 m. Because of the proximities, we assume the two sites within each pair share similar atmospheric conditions. Although the meteorological measurements at the paired sites are not identical, Chen et al (2018) have demonstrated that the differences in meteorology are generally small and not likely a major factor in simulated surface flux differences in most of the pairs. Therefore, the difference (forest vs non-forest) in surface fluxes can be considered as the effects of deforestation.
Sensible (H) and latent (LE) heat fluxes (daily averaged fluxes from 10am to 4pm) are used to calculate the EF (EF = LE/(LE + H)) for both forested and non-forested sites, which serves as a target variable in the following transfer entropy analysis. The canopy photosynthesis flux (gross primary productivity (GPP)) is derived from the measured net ecosystem exchange (NEE) of CO 2 and total ecosystem respiration, the latter was extrapolated from the nighttime NEE using environmental variables (Reichstein et al 2005). We apply GPP as a proxy for vegetation dynamics and stomatal conductance (Williams and Torn 2015). In this study, we also include other available observations (from FLUXNET) in our analysis: downwelling shortwave radiation (R), temperature (T), precipitation (P), and VPD to investigate how atmospheric conditions affect EF. The analysis is conducted at relatively dry and wet sites, which are selected by their annual precipitation being either lower or higher than the mean of all the sites.

Transfer entropy analysis
To quantify directional control of a source variable (e.g. GPP) on a target variable (e.g. EF), information theory states that the uncertainty underlying the target variable could be significantly reduced, if it shares a significant amount of information entropy with the source variable, when excluding effects from confounders (Ruddell and Kumar 2009). Transfer entropy measures the amount of information from a source variable to a target variable. If significant transfer entropy exists, a causal link from a source variable to a target variable exists. Transfer entropy is based on Shannon information entropy (Shannon 1949), which measures the uncertainty of a system, and can be derived from the probability distribution of the variable: where x t is time series of variable X. To estimate the probabilities in equation (1), each variable is discretized into several fixed-interval histogram bins.
x t , therefore, can also be regarded as values in the time series X for the t th bin. H(X) is measured in bits and hereafter will be referred as information entropy. Transfer entropy measures the shared information entropy between source and target variables by excluding the information entropy from confounders. As increasing the number of confounders will cause high-dimensional computation and statistical reliability issues (Runge 2018, Runge et al 2019, we only consider the immediate history of a target variable as the confounder, which always contributes the most confounding information in many study cases (Ruddell and Kumar 2009). The transfer entropy from a source variable (X) to a target variable (Y), knowing the historical condition of target variable, is defined as (Schreiber 2000): where l and k are leading time of source variable X and the historical condition of Y, respectively and y t are values in the Y time series for the t th bin. In the experiments, we use daily data, and set k as 1 day (Ruddell and Kumar 2009). For the time lag, l, 0-5 days are considered, because we focus on short-term processes within a few days (Scott et al 1997, Brunel et al 2006, Ivanov et al 2008. For the binning parameter, we use 11 bins, according to the recommendation of the proper bin numbers from Ruddell and Kumar (2009).
The shuffled surrogate method (Kantz and Schürmann 1996) is employed to determine whether an information flow between coupling variables is significantly stronger than that between shuffled source and target time series by random chance. X s and Y s are obtained by shuffling X t and Y t randomly in time to destroy time correlations. Surrogate transfer entropy T s (X s → Y s ) is computed 100 times using Monte Carlo simulations. A one-tailed significance test is applied to determine the 95% confidence of the transfer entropy (Ruddell and Kumar 2009). The detailed key steps of significant test can be seen in text S1. Finally, relative transfer entropy (T/H) is calculated to normalize transfer entropy with its total uncertainty (Bossomaier et al 2016) and make the strengths of transfer entropy inferred from observations and E3SM land model (ELM) model simulations be comparable.
For each couple of variables (e.g. T/H(X→Y)), we chose the maximum value of causal strength among all the time lags (0-5 days) to represent the strongest interactions within the 5-day time window. Comparison of the strongest causal links within a time window is also widely employed in other causality detection researches (Ruddell andKumar 2009, Ye et al 2015). To make sure all the selected causal links are significant, we conducted significant tests on each potential causal link at each time lag. For the potential causal links that failed to pass the significant test, their transfer entropy was set to zero. In this study, all the analysis is based on the time lags with peak strengths at which most information is transferred (Ruddell and Kumar 2009). We append details of the causal strengths at different time lags in the supplement (text S2, figure S4, tables S3 and S4).

Land surface model simulations
We use offline simulations of the land model (ELM) in the Energy, Exascale, and Earth System Model (E3SM) (Zhu et al , 2020 forced with in situ climate at the paired sites to assess how well the observed patterns of surface energy partitioning are reproduced. Although the offline ELM does not simulate two-way feedbacks between the atmosphere and land surface, it is able to capture one-way causal links from the atmosphere to the land surface. Therefore, the ELM simulated causal controls from environmental factors to EF can be directly compared with those derived from observations. The ELM satellite phenology (SP) mode uses prescribed vegetation phenology (leaf area index) from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite and observed temperature, precipitation, downwelling shortwave and longwave radiation, and vapor-pressure deficit to prognostically simulate ecosystem carbon, water, and energy dynamics. The longitude and latitude of each site are used to extract necessary vegetation and soil properties from a global surface dataset (Koven et al 2013), including e.g.

How deforestation changes the land-surface energy partitioning
First, we investigate the deforestation effects on landsurface energy partitioning between H and LE using the EF. Shifting from forested to non-forested open lands, we did not identify any consistent changes in observed EF (figure 2(a), table S7). In general, deforestation led to significantly decline in LE, H, and R (figures 2(c), (d) and (e), table S7), resulting in relatively small EF effects (table S7). Diurnal and seasonal dynamics of LE and H confirm that the decline of LE and H mainly occur during daytime and during summer, due to limited evapotranspiration and smaller net radiation inputs over non-forested surfaces (Chen et al 2018). Consequently, LE and H offset each other to either enhance or reduce EF ( figure 2(a)). For the non-forested sites, the 25 th , 50 th , and 75 th percentiles of EF were 0.37, 0.46, 0.59, respectively. Compared with the 25 th , 50 th , and 75 th percentiles of EF at forested sites (0.27, 0.48, 0.75), deforestation tended to enhance EF when the forested EF was low, but, conversely, reduced EF when the forested EF was high (detailed pairwise comparison of EF in forested and non-forested sites can be seen in figure S1). Similarly, across precipitation gradients from 0 to 1200 mm yr −1 (figure 2(b)), the differences between forested and non-forested EF at the sites with precipitation from 600 to 1200 mm yr −1 was significantly lower than those at the sites with precipitation lower than 600 mm yr −1 (figure 2(b), text S3). It seems that the EF difference increased again when the mean precipitation higher than 1200 mm yr −1 ( figure 2(b)), but no enough data was available to test this hypothesis.
In general, ELM better reproduced observed EF at the forested sites than non-forested sites as the mean absolute error at forested sites was significantly lower than that of non-forest sites (table S5). Meanwhile, ELM tended to significantly underestimate EF at nonforested sites (figure 3(a) and table S6). Therefore, ELM better reproduced observed EF at the forested sites than non-forested sites. This inaccuracy was due to an underestimation of evapotranspiration and latent heat fluxes at some C3 grass and temperate shrub sites, particularly during spring and summer, as we identified in a recent study (Cai et al 2019). Similar to the observed forested and non-forested EF relationship ( figure 2(a)), ELM also did not exhibit a systematic shift in surface-energy partitioning due to forest cover change ( figure 3(b)). But the ELM failed to represent the systematic shift of LE and H caused by deforestation (figures 3(c) and (d), table S8).

How deforestation affects land part of land-atmosphere coupling
Because of limited observations of soil moisture in the FLUXNET paired dataset, an investigation of the classic soil moisture-EF relationship (Dirmeyer 2011) is not considered here. Instead, vegetation controls over land-atmosphere coupling are evaluated by the relative transfer entropy from Gross Primary Production (GPP) to EF, taking GPP as a proxy for stomatal conductance that regulates water flow from soil to the atmosphere (Baldocchi et al 2001). Such an assumption is largely supported by the close relationship between observed GPP and evapotranspiration fluxes across a wide range of vegetation covers (Mu et al 2007, Gitelson et al 2014, Jiang and Ryu 2016. The observed directional control of GPP on EF at forested sites was significantly higher than that at non-forested sites ( figure 4(a), table S9), indicating a stronger vegetation impact on atmospheric conditions over forests through recycling and convective cloud development (Blyth et al 1994, Trenberth 1999, Freedman et al 2001. However, no statistically significant difference was identified in ELM between forested and non-forested sites in terms of the directional control of GPP on EF ( figure 4(b), table S9). ELM generally captured a similar directional control from GPP on EF at the non-forested sites (figures 4(a) and (b), dashed lines, table S10), but failed at the forested sites (figures 4(a) and (b), solid lines, table S10). Moreover, the results (figures 4(c) and (d), table S12) showed that at relatively dry sites (mean annual precipitation <373 mm yr −1 , figure S2), the observed control from GPP on EF remained similar with that at relatively wet sites (mean annual precipitation >373 mm yr −1 , figure S2). It demonstrates the important role of stomatal controls on ecosystem water loss through evapotranspiration in both relatively dryer and wetter conditions. Similar conclusions have been drawn from studies at U.S. Southern Great Plains sites, where the strong relationship between GPP and EF was observed during both relatively dry and wet periods (Williams and Torn 2015). However, the ELM significantly underestimated the control from GPP on EF at forested sites over relatively dry areas (figure 4(c), table S11). The underestimation was not significant at the forested sites over relative wet areas, as well as at the non-forested sites in both relatively dry and wet areas (figures 4(c) and (d), table S11).

How deforestation affects atmosphere part of land-atmosphere coupling
Together with the controls from vegetation and soil, atmospheric conditions (i.e. temperature and radiation) are also expected to mediate surface-energy partitioning, and consequently contribute to land-atmosphere coupling (Farah et al 2004, Peng et al 2018. For example, when the surfacewater supply is abundant, higher solar radiation (a major source of surface energy) will potentially provide more energy to evaporate surface water and also affect plant stomatal behavior (Pieruschka et al 2010), exerting strong controls on latent heat flux (figure S3). Such controls, however, might weaken during dry periods, especially when the canopy is dense and plants actively close stomata to prevent water loses. In this case, removing vegetation cover might help couple the incoming radiation back again with soil evaporation and EF, even though the soil is relatively dry (Gentine et al 2007).
In the study, solar radiation (R) exerted a much stronger control on EF at forested sites (figure 5(a), table S13), with a mean causal strength 27%-44% larger than the controls from other drivers including temperature (T), precipitation (P), and VPD (figure 5(a), table S13). Deforestation tended to reduce the controls from R, and P on EF, and maintained the directional T, and VPD on EF (figure 5(a), table S14).
ELM model results revealed that although EF values at forested sites were relatively well simulated, the directional controls from R, P, T and VPD on EF were consistently underestimated (figures 5(a) versus (b), table S15). Moreover, the ELM-simulated land segment of land-atmosphere coupling was not sensitive to deforestation or changes in vegetation cover (table S16). The major discrepancy is in how available energy (solar radiation) reshapes the surface-energy partitioning before and after the forest is removed. Figure 4. Land segment of the land-atmosphere coupling. Observed (a) and model simulated (b) probability densities of relative transfer entropy from GPP (proxy for land stomata conductance) to EF for forested sites (black solid line) and non-forested sites (black dash line). The observed (white) and model simulated (dark grey) relative transfer entropy from GPP to EF for forested and non-forested sites in relatively dry sites (c) and relatively wet sites (d).

Implications for land-model development and analysis
Land-surface models have been used to analyze responses of land carbon, water, and energy dynamics to ongoing environmental changes (Zhu and Zhuang 2013, Riley et al 2018, Fleischer et al 2019, Medvigy et al 2019. LULCCs, such as deforestation, significantly disturb the vegetation cover and modify land-surface characteristics (Gash andNobre 1997, Davin andde Noblet-ducoudré 2010), thus imposing critical challenges in evaluating land-surface model fidelity in terms of simulating carbon, water, and energy fluxes before and after deforestation (Lawrence et al 2016). Significant effort has been conducted to parameterize landsurface models and reduce the model-data discrepancy (Decharme 2007, Chen et al 2010, Domínguez et al 2010, Zhu et al 2016, Meier et al 2018, Cai et al 2019, Song et al 2020. However, it has not been fully investigated why model biases exist. This study expands benchmarking of land-surface energyflux partitioning to include cause-effect relationships that might be responsible for model biases. We highlight potential model biases of leaf stomata conductance controls on surface-energy partitioning in the ELM model, especially when the land surface is forested, and the model's significant underestimation of how incoming radiation and precipitation control surface-energy partitioning for both non-forested and forested land surfaces. In contrast, temperature and VPD are much more accurately represented in ELM in terms of their directional control on surface energy partitioning. Therefore, this analysis could direct future ELM model development to particularly focus on vegetation stomata conductance parameterization, radiation transfer, and precipitation partitioning (i.e. intercept, overflow) processes.
Analysis of deforestation effects on surface-energy partitioning is often complicated by compounding changes in vegetation cover and background climate forcings (Winckler et al 2017). Isolating the effects of these changes is challenging in observations but is valuable for improving understanding of how and why deforestation reshapes land-surface energy partitioning. This analysis applied a FLUXNET dataset in which non-forested and forested sites are paired with similar atmosphere conditions (e.g. temperature, precipitation) (Chen et al 2018). Thus, the diagnosed biases in the directional controls could be attributed to the processes that govern the directional controls rather than originating from differentness in background climate forcings.

Conclusion
Deforestation significantly modifies the land surface by changing vegetation cover, disturbing surface-energy partitioning, and land-atmosphere interactions. In this study, we investigated the changes in land-surface energy partitioning with a FLUXNET dataset of paired forested and non-forested sites. We applied nonlinear diagnostic metrics to quantify how land and atmospheric states control surfaceenergy partitioning using transfer entropy calculations. We found that: (a) deforestation simultaneously reduces latent and sensible heat fluxes due to a reduction of both evapotranspiration and incoming radiation, thus possibly leading to either an increase or decrease in EF; (b) GPP exerted a strong control on the EF in forests and deforestation weakens this control; (c) incoming radiation was the most prominent atmospheric variable mediating the EF; (d) the ELM simulated the EF better at forested sites than non-forested sites; however, it failed to capture most of the directional controls from either land segment or atmosphere segment on surface-energy partitioning at both forested and non-forested sites. This study therefore highlights the need for benchmarking model-simulated directional controls on land-atmosphere water and energy fluxes to improve land-model performance.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.