Modeling the effects of realistic land cover changes on land surface temperatures over China

Abstract

In recent decades, China has experienced intense land use and cover change (LUCC). In this study, we modeled and explored the possible effects and mechanisms of realistic LUCCs on surface temperatures in China during 1984–2013 using the Regional Climate Model (RegCM) based on the annual dynamic LUCCs derived from Global Land Surface Satellite (GLASS-GLC) data. We compared two sets of ensemble experiments, one with a fixed LUCC scenario (fixed at 1984) and the other with a dynamic LUCC scenario (annual conversions from 1984 to 2013). The results showed that LUCCs in recent decades have been characterized by reforestation (accompanied by a reduction in cropland) in southern China, grassland-to-cropland conversions in North and Northeast China and bare land-to-grassland conversions in northwestern China. Such LUCCs led primarily to significant cooling of the daily maximum surface temperature (T_smax) by approximately − 0.3 to − 0.6 °C in summer and autumn over the reforestation areas of southern China, followed by moderate cooling (~ − 0.2 °C) of the daily mean surface temperature (T_s), while the effect on the daily minimum surface temperature (T_smin) was weak (within ± 0.1 °C) and nonsignificant. During winter and spring, the impacts of LUCC on all temperatures were less pronounced. Further investigation revealed that the cooling of T_smax was dominated by the direct effects of LUCC (over 90%), particularly the reduction in aerodynamic resistance associated with the cropland-to-forest conversions in southern China. Additionally, seasonal differences and uncertainties in the LUCC-induced cooling of T_smax were likely due to a combination of the relatively stable direct effect of LUCC and the uncertain model internal variability. Overall, our study highlights the effects of realistic and transient LUCC scenarios, which could provide insightful scientific clues to better assess and understand climate change.

Appendix: preprocessing steps of the GLASS-GLC data

1.1 Converting from the original land cover types to model resolution fractions

First, we aggregate the land cover types at a 5 km resolution to land cover fractions at the model resolution (i.e., 25 km × 25 km) by dividing the number of 5 km pixels by the total pixels falling in each model grid. After this step, the original land cover types of GLASS-GLC are transformed into fractional information (such as the PFT fraction in CLM4.5) for each model grid.

1.2 Smoothing and calculating anomalies

Next, we analyze the main spatio-temporal LUCC patterns in the GLASS-GLC data during 1982–2015 using the fractional data from step 1, by applying multiple statistical methods (i.e., empirical orthogonal function method and linear trend analysis; Figs. S1 and S2). The results show that four land cover categories (i.e., bare ground, forest, grassland and cropland) have experienced robust changes in both space and time. In terms of corresponding spatial patterns, the LUCCs during 1982–2015 exhibit conversions from cropland to forest in southern China, from grassland to forest in North and Northeast China and from bare ground to grassland in northwestern China (Figs. S1 and S2). Furthermore, the four major classes of LUCC patterns in China from the GLASS-GLC data generally show consistent near-linear trends in their temporal evolution (Figs. S1, S2). Notably, the spatial and temporal variabilities of the three remaining land cover types in GLASS-GLC (i.e., shrubland, tundra and snow/ice) did not change or were not present during 1982–2015 in China. Therefore, when processing the data of shrubland, tundra and snow/ice, the percentages are maintained and fixed to the values in the default land cover data (i.e., Lawrence and Chase 2007) for each year during 1982–2015.

Moreover, two additional processes are applied to further reduce the potential uncertainties in the LUCC information obtained from the GLASS-GLC dataset based on the above EOF results. First, in order to minimize the effect of possible spurious or erroneous inter-annual fluctuations in satellite time series data on the modeling results, we apply a 5-year running mean to the original fractional information from step (i). The main reasons are as follows: (1) The interannual fluctuations in satellite data can be potentially affected by various aspects (e.g., atmospheric variability and aerosols; Xiao et al. 2003; Lu et al. 2007), and thus their interannual variability is relatively more uncertain; (2) The effects of interannual fluctuations in producing land cover data cannot be completely avoided (Song et al. 2018; Liu et al. 2020), hence, it is therefore difficult to distinguish whether the interannual fluctuations reflected in the data are signals or noise; (3) We observed from the statistical analysis above that the first-order (absolutely dominant) modes of LUCCs in China were close to linear trends, so we filter out the relative negligible high-frequency interannual fluctuations in order to focus more on the dominant LUCC variability in the GLASS-GLC data. Hence, after the 5-year running mean process, the high-frequency interannual variations in the LUCC information can mostly be screened out, and the smoothed results represent the first-order features (i.e., near-linear trends) of LUCC information in the GLASS-GLC data.

Second, we chose to use the “static basemap plus annual anomalies” approach, rather than using direct replacement of annual land cover maps, to generate the annual land cover maps was based on the following considerations. Firstly, both the “basemap plus anomalies” approach and the direct replacement of annual land cover maps are identical for extracting the dynamic LUCC information from GLASS-GLC data. Secondly, the direct replacement and production of annual land cover maps with new data is equivalent to the introduction of both GLASS-GLC basemap (e.g., the map of 2000) and annual anomalies; the caveat here is the considerations of the basemap replacement. It is known that GLASS-GLC and the model’s default land cover basemap (based on MODIS data) are two very different datasets, which differ significantly in terms of raw data sources, production methods and classification systems. If GLASS-GLC data are used to completely replace the model’s default (i.e., MODIS) land cover data, mismatches may occur between the land cover types and consequently their corresponding vegetation parameters, which in turn may introduce potential uncertainties. Furthermore, the critical information we are more interested in is the anomalies information of LUCC, and replacing the basemap may perturb the entire matrix of the model (e.g., the choices of parameterization schemes and the settings of parameters) to the extent that the performance of the model after replacing the basemap may need to be re-evaluated from scratch. This is far beyond the scope of our current study. Therefore, we chose to superimpose the GLASS-GLC PFT anomalies onto the model’s default MODIS PFT data rather than using a full replacement approach to generate the annual transient land cover maps required by the model. This mapping is completed as follows. (1) The year 2000 of the 5-year running average GLASS-GLC PFT fraction data is used as the baseline, which represents the 5-year average PFT state during 1998–2002. (2) The baseline PFT fractions (i.e., the year 2000) are subtracted from the 5-year running average PFT fractions for each year during 1984–2013 to obtain the GLASS-GLC PFT anomalies spanning 1984–2013. (3) The anomalies from the previous step are superimposed onto the model’s default MODIS PFT fractional data for each year. After these steps, annual transient land cover maps based on GLASS-GLC LUCC information are produced for 1984–2013. Notably, we do not process the grids (i.e., we retain the LUCC information in the default PFT input data during the whole period) where the percentage of any type of grid point exceeds 100 or is less than 0 when the anomaly is added, as these grid points are possible sources of uncertainty that may arise due to different land cover data sources (i.e., GLASS-GLC versus MODIS).

1.3 Adapting LUCC information to the model

Among the GLASS-GLC land cover types, cropland and barren land (i.e., bare ground) are the same as or similar to those in the CLM4.5, so their fractional information can be directly applied to the model. The rest, i.e., forest and grassland, cannot be directly utilized due to the lack of subtype information. Therefore, we created additional fractional information for forest and grassland subtypes based on the relative proportions of subtype PFT information in the default PFT dataset of the CLM4.5 (i.e., Lawrence and Chase 2007). This process allows the total PFT for forest and grassland to be variable, while the relative proportions of subtypes remain fixed under the annual transient LUCC scenario. This approach to creating additional information regarding subtypes is also similar to that of Liu et al. (2021).