Afforestation has been and is still proposed as an effective strategy to mitigate climate change because forest ecosystems are able to sequester large amounts of carbon in their biomass and soil, slowing the increase of atmospheric CO2 concentration (Fang et al., 2014; Pan et al., 2011). Additionally, forests regulate the exchange of energy and water between the land surface and the lower atmosphere through various biophysical effects, including radiative processes such as surface reflectance and non-radiative processes such as evapotranspiration and sensible heat flux (Bonan, 2008; Juang et al., 2007). As the net result of the surface energy balance, land surface temperature (LST) is widely used to measure the local climatic impact of afforestation (Li et al., 2015; Winckler et al., 2019a).

Climate model simulations and site-level observations have been utilized to explore the impact of forest dynamics on land surface temperature (Lee et al., 2011; Pitman et al., 2009; Swann et al., 2012). However, afforestation impacts on local LST derived from models tend to be highly uncertain as they are limited by the coarse spatial resolution of models and uncertainties in model parameters and processes (Oleson et al., 2013; Pitman et al., 2011), while insights from site-level assessments cannot be extrapolated to large spatial domains (Lee et al., 2011). Alternatively, remote-sensing-based LST products enable the assessment of local LST changes due to forest dynamics on large spatial scales (Li et al., 2015; Shen et al., 2019).

A number of studies investigated the surface temperature impact of afforestation based on satellite observations, and they have been cited in high-level climate science synthesis reports (e.g., IPCC Special Report on Climate and Land, authored by Jia et al., 2019), even though there are large differences in afforestation impacts on LST between different methods. For example, Alkama and Cescatti (2016) found a cooling effect of about 0.02 K from afforestation in temperate regions, while Li et al. (2015) reported a 0.27±0.03 K potential cooling from afforestation in the northern temperate zone (20–50∘ N) based on the “space-for-time” method. In contrast, Duveiller et al. (2018) found a much stronger potential cooling effect of 2.75 K for afforestation in the northern temperate region. While such differences were acknowledged in a recent modeling study (Winckler et al., 2019b), they were simply treated as observational uncertainty for climate model evaluation, with the uncertainty range being as big as, or even an order of magnitude larger than, the afforestation effect. It remains unclear whether the differences arising from these different methods can be reconciled.

Until now, studies using satellite data to investigate afforestation impact on surface temperature have mainly focused on three methods. The first method, termed the “space-and-time” approach (Fig. 1, red box), aims to examine the actual, realized effect of afforestation (“actual effect”) by isolating the forest-cover-change effect from the gross temperature change over time in places where forest-cover change actually occurred (Alkama and Cescatti, 2016; Li et al., 2016a). The second method, termed the space-for-time approach (Fig. 1, orange box), compares the surface temperature of forest with adjacent “openland” (i.e., cropland or grassland) under similar environmental conditions (e.g., background climate and topography) and estimates the potential effect of afforestation if afforestation were to occur (Ge et al., 2019; Li et al., 2015; Peng et al., 2014). Note that such effects are often quantified using medium-resolution land-cover datasets (typical resolution = 1 km), which do not necessarily represent 100 % ground coverage, and we therefore term such a potential effect a “mixed potential effect”.

The third method, recently used by Duveiller et al. (2018), uses the “singular value decomposition” (SVD) technique (Fig. 1, green box), which is claimed to extract the hypothetical LST for different land-cover types by assuming a 100 % coverage of the target cover type. The afforestation effect on LST is then quantified as the difference between the LST of a pixel with a hypothetical 100 % forest coverage and the LST of an adjacent pixel with 100 % openland coverage. As with the second method, such an approach quantifies the potential effect of afforestation, but in this case, it quantifies the “full potential effect” by assuming transitions between land-cover types with 100 % complete ground coverage.

Previous studies have revealed the fraction of forest change as an important factor determining the magnitude of the afforestation effect. Alkama and Cescatti (2016) indicated that the actual temperature effect is fraction-dependent, and Li et al. (2016a) pointed out that use of a higher threshold to define forest change resulted in a stronger potential effect. Nonetheless, whether the fraction of forest change can explain the differences in the afforestation effect produced by different methods, e.g., whether the potential effect can be “actualized”, has not been demonstrated. Testing the role of afforestation fraction in reconciling the afforestation effects produced by different methods can help clarify potential confusion and contribute to appropriate policy evaluation.

This study develops detailed conceptual and methodological descriptions for each of the three approaches and uses large-scale afforestation over China as a case study to compare the three approaches. We tested the following hypotheses. (1) The actual effect on LST increases with the area that has actually been afforested, defined as afforestation intensity (or Faff). (2) The actual effect is smaller than the potential effects. (3) When extending Faff to a hypothetical value of 100 %, the actual effect approaches the potential effect. If proven, this third hypothesis implies that the LST impacts from different approaches could be reconciled given the same boundary condition of full afforestation. In that case, we then have a fourth hypothesis (4) stating that changes in underlying biophysical processes including radiation and sensible and latent heat fluxes that drive LST changes should also be reconciled among different methods. To keep the focus on reconciling methodological differences, only changes in the daytime surface temperature were considered in this study. Nevertheless, similar conclusions regarding the different approaches are expected for nighttime surface temperature.

The three approaches (Li et al., 2015; Alkama and Cescatti, 2016; Duveiller et al., 2018) used to quantify local surface temperature change following forest-cover change and presented in detail in this study have been cited over 919 times in research papers (Web of Science, December 2021) and in high-level climate science synthesis reports. Despite the apparently large differences in temperature effect among them, to our knowledge, no studies have examined whether these differences can be reconciled. This study fills that gap by comparing the three approaches for a single study case, i.e., large-scale afforestation in China. China is highly suitable for the purpose of this study as the size of an afforestation patch is, in general, smaller than the spatial resolution (1 km) at which the temperature effects of afforestation were conducted in the previous studies describing the three approaches (Li et al., 2015; Alkama and Cescatti, 2016; Duveiller et al., 2018). Hence, the difference between the actual and potential temperature effects is expected to be large.

Indeed, we found surface cooling following afforestation was much less when estimated as the actual effect (ΔTa) compared to the potential effects (ΔTp1 and ΔTp2). This lower ΔTa has been attributed to incomplete afforestation at a 1 km resolution, at which potential effects are quantified by assuming complete afforestation (i.e., a complete shift from openland to forest). Consistent with our first hypothesis, the afforestation fraction at a 1 km resolution explained 89 % of the variation in ΔTa, making it a key determinant of the surface cooling following afforestation (Fig. 5). This result is consistent with the fact that the observed temperature for a mixed surface is determined by the area fractions of its respective components, with each component having a unique temperature. This fact also forms the theoretical foundation for the SVD technique used to derive the full potential effect (Duveiller et al., 2018).

Modeling (Li et al., 2016b) and satellite-based (Alkama and Cescatti, 2016) studies have found that temperature change after afforestation (or deforestation) is highly sensitive to the fraction of the model grid cell or satellite pixel that is subjected to afforestation (or deforestation), echoing our finding that ΔTa significantly changes with Faff. In addition, we provide strong evidence in support of our third hypothesis that when Faff reaches 100 %, the expected actual effect is comparable to the potential effects (Fig. 7). This finding shows that the three approaches compared here are consistent when the same boundary condition, i.e., full afforestation, is applied and demonstrates that all three methods are mutually compatible. It is, therefore, the basis of the reconciliation of the three approaches. It also highlights the fact that the actual afforestation area must be considered when evaluating the climate mitigation effects of afforestation.

Our results also show that the mixed potential effect (ΔTp1) is smaller than the full potential effect (ΔTp2) (Figs. 6, 7). We suspect that this phenomenon likely also relates to the incomplete forest coverage for the identified forest pixels at the 1 km scale used in the space-for-time analysis because a threshold value of 50 % forest cover was used when upscaling the 30 m land-cover map to 1 km resolution. This threshold, however, is consistent with the commonly applied value in land-cover classification based on medium-resolution satellite images, such as MCD12Q1, which uses a tree coverage value of 60 % to identify forest pixels (Sulla-Menashe et al., 2019). For the purpose of comparison, we also calculated the mixed potential effect based on the MCD12Q1 land-cover map but using the same LST data. The result shows that potential effects derived using MCD12Q1 data versus those derived using spatially upscaled GlobeLand30 data are almost identical (Fig. A5), lending credibility to our estimated ΔTp1 in comparison to previous studies using MODIS land-cover data (Li et al., 2015). Progressively increasing the forest-cover threshold from 50 % to 90 % steadily increases ΔTp1 from -0.62±0.02 to -0.75±0.02 K (Fig. A6). Further increasing the thresholds used to identify 1 km resolution openland pixels from 50 % to 90 % increases ΔTp1 from -0.63±0.00 to -1.10±0.02 K (Fig. A7), bringing ΔTp1 even closer to ΔTp2 (-1.16±0.01 K). This is consistent with the finding of a previous study on the dependence of the temperature effect on the forest-cover-change thresholds that were used to define afforestation: the higher the threshold, the stronger the impact on temperature (Li et al., 2016). Our results add further support to the compatibility of the three approaches given the same boundary condition, i.e., the complete transformation from full openland to full forest coverage.

Several factors may contribute to the remaining differences in the temperature effects produced by different methods even after reconciliation. As described in the Method section, there are discrepancies in the assumptions of the three approaches, which lead to differences in the control pixels (i.e., adjacent comparison pixels). For instance, for the “pure” potential effect, it is assumed that the LSTs for pixels with the same land-cover type are uniform, and forest pixels are compared with openland pixels, whereas in the mixed potential effect approach, the central target forest pixel is compared with the mean value of non-forest pixels within the searching window. Moreover, space-for-time is an assumption that cannot be verified (Chen and Dirmeyer, 2016) and which will inevitably result in differences in the estimated actual and potential effects, although the consistency between potential and “actual” effects after reconciliation in our study does illustrate the broad validity of this assumption.

Differences between the actual and potential temperature effects can also arise from influences of both the timing of the afforestation and the time period elapsed following afforestation. However, such influences are expected to be small in our study. We argue that such influences should be more pronounced in the case of deforestation than afforestation. The temperature effect caused by deforestation is considered to be instant (Liu et al., 2018). As a result, if deforestation occurred in 1 specific year of our starting time window (i.e., 2002–2004), using the time-averaging LST over the whole time window to represent the LST before deforestation will greatly bias the quantified ΔT. In contrast, afforestation-driven surface temperature change can only gradually increase with forest development. The LST effect depends on different stages of forest development and is expected to saturate only when the forest canopy stabilizes (Zhang et al., 2021; Windisch et al., 2021). Observation studies show that closed dense-canopy old forests can exert a greater cooling effect than the open-canopy young forests (Zhang et al., 2021; Windisch et al., 2021; Duveiller et al., 2018, 2020). Hence, given the gradual nature of the afforestation effect on LST, when we quantify the afforestation effect by comparing the time-averaging LST before and after afforestation, the influence of the specific timing of afforestation is expected to be small. Furthermore, the GFC dataset used in this analysis defined forest gain using the condition of successful detection of a stable closed forest canopy that is clearly different from a non-forest state (Hansen et al., 2013), which enhances the chance of temperature change saturation following afforestation. But, given a maximum stand age of 12 years inferred from the GFC dataset, differences in surface temperatures may still exist between newly established forests and the mature existing forests that were used in the potential effect approaches. Thus, we cannot exclude the possible contribution of time period elapsed following afforestation to the difference between the actual and potential effects, which failed to be reconciled.

Previous analyses have documented latitudinal patterns of surface temperature change induced by afforestation (Alkama and Cescatti, 2016; Li et al., 2015, 2016a; Peng et al., 2014). When comparing the three approaches for a single case study, consistent latitudinal patterns of local surface temperature effects following afforestation are observed (Fig. 4). Notably, all three approaches show a warming effect in the northern high latitudes and an opposite cooling effect in the southern low latitudes, with a largely neutral effect in the 40–48∘ N latitude band, providing further evidence that the three approaches are compatible. In particular, although the three approaches used different land-cover maps, they derived consistent LST impacts following afforestation, which highlights that fact that the reconciliation provided in this study is rather robust and unlikely to be dependent on the land-cover datasets used.

In addition to the reconciliation of the land surface temperature change, we checked and confirmed that the changes in surface energy fluxes that underlie and drive the changes in surface temperature are compatible under the boundary condition of full afforestation. This finding confirms the inherent consistency in the three approaches and clarifies the reasons behind the apparent discrepancies in existing studies as discussed in the Introduction. Nonetheless, when it comes to the biophysical impacts of afforestation in the real world, our findings have far-reaching implications. Full afforestation is often possible at small spatial scales but becomes challenging at large scale. Therefore, the realization of the full potential effect by afforestation is scale-dependent. For example, a complete afforestation of the semi-arid Loess Plateau in the northwest of China is predicted to generate a surface cooling effect of 2.40±0.07 K, but substantial afforestation efforts over the past 4 decades in that region have only realized a cooling of 0.11±0.01 K, as measured by the actual effect. Because of greater water consumption by forest compared to openland and the need to maintain land area for food production, achieving the full cooling potential may not be feasible (Huang et al., 2018; Liu and She, 2012; Liang et al., 2019).

Potential cooling effects have a value in that they can serve to establish the envelope of effects and measure possible outcomes given the condition of full afforestation. However, given the challenge of full afforestation at large spatial scales, potential effects should be converted into a more realistic estimate (i.e., actual effects), by taking into account the intensity of afforestation, to better represent policy ambitions. The analog could also be made for the effects of the surface energy impacts of afforestation. Taking 10 % as the afforestation intensity threshold to compare the cumulative surface energy effect between the actual and potential approaches, actual cumulative biophysical changes (5.06 EJ) for 2000–2012 are much smaller than mixed potential changes (20.13 EJ) and full potential change (19.02 EJ) (methods in Sect. A2 in Appendix A; Fig. A8). Again, this shows that simply using the potential effects for policy making or evaluation risks greatly overestimating the biophysical effects of afforestation.

In this study we provided a synthesis of the three influential methods used to quantify afforestation impact on surface temperature change and provided evidence that these different methods could in fact be reconciled. The actual effect of surface temperature change following afforestation was highly dependent on the intensity of afforestation (Faff), which explained 89 % of the variation in ΔTa. With the common boundary condition of full afforestation being applied, differences in afforestation impacts on LST reported by the three methods in previous studies greatly reduced, showing that simply treating these differences as uncertainty is incorrect and could greatly overestimate the uncertainty. In other words, when full afforestation is assumed, the actual effect approaches the potential effect, demonstrating the effectiveness of the space-for-time approach and that the potential cooling effect of afforestation could indeed be realized. Potential cooling effects have a value in academic studies where they can be used to establish an envelope of effects, but their realization at large scales is challenging given the scale dependency. The reconciliation of the different approaches demonstrated here stresses that the afforestation fraction should be accounted for in order to bridge different estimates of surface cooling effects in policy evaluation.