Satellite evidence of canopy-height dependence of forest drought resistance in southwestern China

The frequency and intensity of drought events are increasing with warming climate, which has resulted in worldwide forest mortality. Previous studies have reached a general consensus on the size-dependency of forest resistance to drought, but further understanding at a local scale remains ambiguous with conflicting evidence. In this study, we assessed the impact of canopy height on forest drought resistance in the broadleaf deciduous forest of southwestern China for the 2010 extreme drought event using linear regression and a random forest (RF) model. Drought condition was quantified with standardized precipitation evapotranspiration index (SPEI) and drought resistance was measured with the ratio of normalized difference vegetation index during (i.e. 2010) and before (i.e. 2009) the drought. At the regional scale we found that (a) drought resistance of taller canopies (30 m and up) declined drastically more than that of canopies with lower height under extreme drought (SPEI < −2); (b) RF model showed that the importance of canopy height increased from 17.08% to 20.05% with the increase of drought intensities from no drought to extreme drought. Our results suggest that canopy structure plays a significant role in forest resistance to extreme drought, which has a broad range of implications in forest modeling and resource management.


Introduction
Forests are one of the most important natural resources in the world. They provide extensive ecosystem services ranging from producing food, fuel, shelter, medicine and fresh air (Hendrey et al 1999, Carle andHolmgren 2008) to watershed protection, erosion prevention and climate mitigation (Patric 1976, Lu et al 2001, Canadell and Raupach 2008. Several lines of evidence indicate that the frequency, intensity, and duration of land drought events have been increasing along with global climate warming (Yi et al 2014, Toby 2020, which have substantial impacts on the structure and function of forest ecosystems. Insufficient water during drought can affect the material transport in trunk conduits of trees, which may result in the damage of conductive vessels (Delzon andLoustau 2005, Niu et al 2014). Severe drought may cause widespread tree mortality and dieback, and even forest degradation across the region (Huang et al 2015, Kolb 2015. Consequently, the impact of drought can be assessed with the changes of forest structure and functions. The impact of drought on forests essentially depends on which part of the forests are more sensitive to water deficit (Bennett et al 2015). Greater impacts on short and small trees may modify future forest succession whereas greater impacts on tall and large trees causes disproportionate losses of carbon reserve (Phillips et al 2010, Lindenmayer et al 2012. Previous studies have reached a general consensus on the size-dependency of forest resistance to drought (Francesco et al 2018, Xu et al 2018a, 2018b, 2018c. But further understanding at the mechanistic level remains ambiguous and controversial. It is unclear whether large (tall) or small (short) trees would suffer more under drought stress. Some studies indicated that small and short trees suffered more because of their shallow roots and less available water in the deep soil (Nakagawa et al 2000, Guarín and Taylor 2005, Zhang et al 2017, Francesco et al 2018, Liu et al 2022, while other studies indicated that drought had a greater impact on large and tall trees because they have a greater evapotranspiration rate and higher water demand (Aber et al 2001, Nepstad et al 2007, Zhang et al 2009, Xu et al 2018a, 2018b, 2018c. Both views were mainly based on the results from field measurements and control experiments at local scale or site level. These conclusions are greatly affected by site specificity and sample size, and it is challenging to extrapolate the pattern at a larger scale. Since canopy height and structure can be readily changed by human activities, a solid understanding on sizedependent forest resistance to drought at regional scale would provide valuable scientific guidance for policy and practice of forest management (Bellassen and Luyssaert 2014).
As an efficient data acquisition tool, remote sensing observation can provide high spatial-temporal resolution images for forestry related research at a large scale with continuity. Various vegetation indices are based on the contrast that green leaves strongly absorb radiation in the red wavelengths and reflect radiation in the near-infrared (Assal et al 2016). Previous studies showed those vegetation indices are sensitive to various biophysical parameters (such as leaf area index, crown closure and water content of leaves) that could be used to indicate vegetation growth status at a large scale (Deshayes et al 2006). Normalized difference vegetation index (NDVI) is one of the most widely used satellite-based algebraic indices allowing cross-site comparation, which leverages different parts of the electromagnetic spectrum for disturbance assessment , which is suitable for monitoring vegetation growth change from mesoscale to large scale (McManus et al 2012, Vicente-Serrano et al 2013. When vegetation declines as a result of climate stress (such as drought and high temperature), NDVI will decrease.
On the contrary, when the environmental condition is favorable for vegetation growth, NDVI will increase. Therefore, NDVI is suitable for quantifying the impact of climate change on vegetation growth (McManus et al 2012, Vicente-Serrano et al 2013, Huang et al 2015. Remote sensing-based VI allows rapid assessment of forest dynamic changes in response to climate stress such as drought at a regional scale, but often overlooks the role of stand structure in forest response to drought (Assal et al 2016).
In this study, we attempted to assess the forest resistance to drought in southwest China. The objective of this study was to quantify the role of canopy height and its relative importance in forest resistance to drought.

Study region
The study area was located in southwestern China, including the Yunnan, Guizhou, Guangxi and Sichuan provinces and Chongqing municipality (figure 1), which historically had ample precipitation and water resources. Periodically, heavy rains brought by EL NIÑO events such as in 1997/1998 and 2015/2016 also impact the region (Fan et al 2017). However, regional droughts have appeared more and more frequently since 1950 due to a combination of climate change, deforestation, land use change and degradation of hydrological facilities (Qiu 2010, Zhang et al 2013, Zhang and Zhou 2015. Southwestern China experienced a severe drought event in 2010, which had a substantial impact on the local forest ecosystems with widespread mortality and dieback. Since the 'Grain for Green' program (GGP, an ambitious conservation program initiated in 1999 that involved 25 provinces in China to restore cultivated land into forest and pasture) was implemented in this area (Xiao 2014), many new trees were planted which complicated the structure of forest canopy. Therefore, with existing high plant biodiversity in this region, southwestern China was an interesting region in which to study the effects of canopy height on forest drought resistance (Xu et al 2018a(Xu et al , 2018b(Xu et al , 2018c. The forest distribution map (figure 1) was based on the MODIS land cover product (MCD12Q1, Version 6) global land cover types at yearly intervals derived from six different classification schemes (https://modis.gsfc.nasa.gov/). Land cover data derived from IGBP (International Geosphere-Biosphere Programme) classification schemes was used in this study, with a spatial resolution of 500 m in 2010. The main forest types in southwestern China included mixed forest, broadleaf deciduous forest (BDF), broadleaf evergreen forest and needleleaf deciduous forest (table 1). Combined with standardized precipitation evapotranspiration index (SPEI) data, we found that 17.04% of the forest suffered  from severe drought in 2010. BDF was the type of forest that was affected most by the 2010 drought in the area, which was chosen as the subject of this study.

Climate data
The SPEI was used to indicate drought intensity and quantify surface water deficit and surplus in this study (Vicente-Serrano et al 2010a, 2010b. SPEI data (figure 2) were obtained from the global SPEI data set, which was based on monthly precipitation and potential evapotranspiration from the Climatic Research Unit of the University of East Anglia (http://sac.csic.es/spei/database.html). This data provided SPEI timescales between 1 and 48 months, with a half degree spatial resolution and a monthly temporal resolution. To correspond to the time scale used for the NDVI data (see section 2.3), SPEI of July 2009 and 2010 with 12 month period was chosen for the analysis (Vicente-Serrano et al 2010b). Based on its SPEI value, drought intensity of a year could be classified into one of the seven categories: extreme drought severe wet (1.5 < SPEI ⩽ 2), extreme wet (SPEI > 2) (McKee et al 1993, Dorman et al 2013. Previous studies have shown that previous year's moisture condition may have a legacy effect on the vegetation growth of current year (Gao et al 2018(Gao et al , 2020. In order to control the confounding effect of previous year's condition, we only included pixels whose SPEI for the previous year (i.e. 2009) was near normal (−1 < SPEI ⩽ 1) in our analysis, which accounted for 89.8% area of BDF in the study area.

Satellite data
2.3.1. NDVI data MODIS-NDVI product (MOD13A3, Version 6) from 2009 to 2010 was used in this study. Vegetation indices data were provided monthly at 1 kilometer (km) spatial resolution in MOD13A3 product (https://modis.gsfc.nasa.gov/). The NDVI average from July to August of the year was calculated to represent forest growth for that growing season (Xu et al 2018b(Xu et al , 2018c).

Canopy height map
The forest canopy height map at the resolution of 1 km (figure 3) was produced by airborne LiDAR canopy height models using the Geoscience Laser Altimeter System aboard ICESat (Ice, Cloud, and land Elevation Satellite) and global ancillary variables (https://landscape.jpl.nasa.gov) (Simard et al 2011). Zhang et al (2014a) verified the accuracy of the canopy height dataset in southwestern China using field survey data. The determination coefficients of the regression of LiDAR-based canopy height over observed tree height from field measurements was

Auxiliary data
The digital elevation model (DEM) data was derived from the Shuttle Radar Topography Mission (SRTM) data of the U.S. Space Shuttle Endeavour (Jarvis et al 2008). The data set was resampled based on the latest SRTM V4.1 data, with a spatial resolution of 1 km. The soil texture data was from Shangguan et al (2013), with a spatial resolution of 1 km. This data describes the composition ratio of clay, sand and silt soil. The vertical profile of soil property was captured All the spatial data projections were converted to WGS84 coordinate system, and resampled to 1 km resolution based on canopy height data by the nearest neighbor algorithm. All regression analyses were conducted in EXCEL (Microsoft Office 2016), and graphs were made in IDL8.5 and Arcgis10.0.

Forest drought resistance
Resistance is considered as reversal of the reduction in ecological performance during disturbance, and it is estimated as the ratio between the performance during and before the disturbance (Kaufman 1982, Lloret et al 2011, Zhang et al 2014b. In this study, forest drought resistance (Rt) corresponds to the ratio between growth rates during the drought and immediately before the drought (Zhang et al 2014b, Yi andJackson 2021

RF model
To quantitatively assess forest size-dependent resistance to drought, we used RF algorithm to analyze the contribution of canopy structure (canopy height), water condition (SPEI), elevation (DEM), soil texture (clay and silt) to forest drought resistance. The RF was an effective algorithm for optimizing learning accuracy without obviously complicating the calculation, and it has been widely used in geography and ecology research (

Impact of drought severity on forest resistance
The mean resistance (±1 standard error) of a BDF with a short canopy (<20 m) was 0.968 ± 0.012 under extreme drought (SPEI < −2.0), 1.002 ± 0.008 under severe drought (−2.0 < SPEI ⩽ −1.5), 0.998 ± 0.007 under moderate drought (−1.5 < SPEI ⩽ −1.0) and 1.000 ± 0.004 with no drought (−1.0 < SPEI ⩽ 1.0) (figure 4(a)).  4(c)). These results suggested that forest resistance to drought declined as drought severity increased, with the least resistance being under the extreme drought condition. Such decline of forest resistance is also canopy size dependent. When canopy height was over 30 m, resistance dropped drastically (as low as 80%). These results indicated that BDF might have sizedependent resistance to drought and the resistance of taller forests was more sensitive to water condition than shorter forests.

Relationship between forest drought resistance and canopy height
Linear regression models were established between forest resistance and canopy height under extreme drought, severe drought, moderate drought, no drought (figure 5, tables S3-S6). Resistance of forests with various canopy heights was less than 1 which indicated that forest growth declined. There was a significant negative correlation between forest drought resistance and canopy height, and forest drought resistance declined significantly with the increase of canopy height under extreme drought condition (p < 0.001, figure 5(a)). When canopy heights were over than 30 m, the resistance decreased dramatically with canopy heights (green line in figure 5(a)). When the canopy height was over 30 meters, forest drought resistance decreases by 0.041 for every meter increase in canopy height, while forest drought resistance only decreased by 0.0037 for every meter increase in canopy height as canopy height less than 30 meters ( figure 5(a)). In contrast, regressions of DBF resistance to drought under less than extreme drought, regardless of canopy heights, were around 1 which indicated that forest growth was not affected. There was a weak linear correlation between forest drought resistance and canopy height, and forest drought resistance increased slightly with the increase of canopy height under less than extreme drought conditions (figures 5(b)-(d)). Those results suggested that forest resistance to extreme drought was size-dependent and taller forests had a weaker resistance to drought than shorter forests. But the relationship between canopy height and forest drought resistance changed with the change of water conditions. In other words, forest drought resistance is only size dependent under extreme drought, the improvement of water conditions could change this dependence.

Relative importance of canopy height to forest drought resistance
RF model for drought resistance in BDFs in southwestern China showed that the importance of elevation was 51.27%-55.76%, and that of soil texture were 16.82%-20.99% for all four drought conditions with no consistent patterns (table S1). However, the importance of canopy height increased steadily from 17.08% to 20.05% with the increase of drought intensities (from no drought to extreme drought condition, with decreasing SPEI) (figure 6). Although the influence of environment variables was dominant, the influence of canopy heights over forest resistance to drought increased with drier conditions.

Forest drought resistance distinguished by canopy height
Our analyses on satellite data have shown forest drought resistance is not only affected by water condition but also depends on canopy structure at the regional scale in southwestern China (figures 4 and 5). In the past decades, some of the results showed that small trees are more vulnerable, while some suggested that drought have a greater impact on large trees (Nakagawa et al 2000, Guarín and Taylor 2005, Zhang et al 2017, Francesco et al 2018. Compared with previous studies, this study addressed the two opposite findings at a much larger scale which is Generally, our results support the hypotheses that at local scale and site level, larger and taller forest would suffer more from growth reduction during drought regardless of the spatial scale employed (Zhang et al 2009, Xu et al 2018c. Recent studies have also shown that compared with variables such as saturated vapor pressure difference, maximum temperature, precipitation, effective water storage, vegetation coverage and slope, tree height is the most powerful variable for predicting tree mortality (Stovall et al 2019). There are three mechanisms potentially contributing to the canopy-heightdependent responses: physiology and metabolism requirement of individuals with different height (Eamus et al 2013, Sevanto et al 2014, competitive or nursing interaction with neighboring trees (Mueller et al 2005, Andivia et al 2018, and accumulative effects and legacy effects of various dimensions of the drought regime (Gao et al 2018(Gao et al , 2020. Firstly, large and tall trees have a greater evapotranspiration rate and higher water demand than small and short trees, a big disadvantage to drought resistance (Zhang et al 2009). For taller trees, they have a longer vertical water conducting passage from root to canopy, requiring a greater pulling force against gravity generated by transpiration. As water deficit decreases the transpiration rate, the pulling force would be insufficient to transport water to the canopy through xylem, leaving taller trees more vulnerable to embolism and desiccation with a lower drought resistance (Cochard et al 2021). Under drought conditions, extensive root systems also grant larger trees a higher likelihood of absorbing water with air. The air would form bubbles in the xylem to block the conduits and affect the material transport, which may result in tree mortality or growth decline. In addition, carbon sequestration and storage increase with age, peak around 22 years then decline . Net primary production increases with stand age before stabilizing while the metabolism of large trees consumes more dry matter, which collectively increases the risk of mortality in large trees (Waggoner and Turner 1971).
Secondly, individuals with different statures may experience different interactions with neighboring trees, ranging from fierce resource competition at the canopy height to being sheltered from heat and drought at the sub-canopy level (Mueller et al 2005, Andivia et al 2018. Competition as large trees are much stronger than small ones due to their higher demand in radiation, water and nutrients.
Thirdly, there are different facets and dimensions of drought regime that may trigger differential responses to trees with different sizes and ages. Drought frequency (number of dry months) and duration (maximum number of consecutive dry months) resulted in 'cumulative effects' which amplified the impacts of drought on trees and reduced the drought resistance of trees. Onset timing and severity of drought increased legacy effects on tree stem radial growth, which reduced the drought resilience of trees (Gao et al 2018). Although taller trees may have a more extensive root system for absorbing soil water, the intensity of extreme drought was high enough to deplete soil moisture at a deeper level like in the case of 2010 drought in southwestern China, greater damage would happen to taller trees (Xiao 2014, Xu et al 2018c.

Implication of canopy height-dependent forest drought resistance
With global warming, the frequency and duration of drought would increase worldwide in the near future reported by IPCC5 (Stocker 2014). Human activities (e.g. deforestation and afforestation) have drastically altered the structure and functions of forests (Xiao 2014). Recent satellite data (2000Recent satellite data ( -2017 reveal a greening pattern that is strikingly prominent in China and India. China alone accounts for 25% of the global net increase in leaf area with only 6.6% of global vegetated area, and 42% of the greening in China is from afforestation (Chen et al 2019). Such human activities have profoundly impacted forest distribution and structures, including age structure and canopy structure. Forest canopy-height-dependent resistance to drought provides a new perspective and evidence on managing forest in the context of climate change. Humans might reduce forest drought-related mortality or dieback by manipulating the forest structure, specifically by removing tall trees and filling the gaps with transplanting or natural regeneration.
Although the economic value and ecosystem functions of forests increase with age and size over time (Blanco 2012), their vulnerability to drought stress also increase with age and size based on this study. Although taller trees may have a disproportionally higher growth rate, it could be due to their greater light exposure (Araujo et al 2020). Based on the results of this study, proper management of forest canopy structure (i.e. removing aging tall trees that higher vulnerability to drought) would expose subcanopy trees to more light and boost their productivity, and also manage the risk of dieback during extreme drought events.
Our results will also shed new light on optimization of process-based ecological modeling. Many computer simulations models of forest, such as Forest Vegetation Simulator (Crookston and Dixon 2005), Bio-Geochemical Cycle model used by the USDA (United States Department of Agriculture), are based on the assumption of equilibrium (Running andGower 1991, Zhou et al 2015). The input and output of energy and nutrients in a system in equilibrium are in dynamic balance with no net gain or loss over time. This does not hold true for most systems except the climax communities. If we consider the stage structure of the community and factor in its spatial heterogeneity and uncertainty when modeling a forest community, the model would be more realistic and provide more accurate predictions.

Uncertainty
The data used in this study include MODIS products, LiDAR inversion data and gridded climate data. The spatial resolution of those data is not exactly the same, so the deviation of data matching is unavoidable. Although NDVI is sensitive to spectral change caused by disturbance for a wide range of canopy coverage, it could get saturated at very high vegetation coverage and lead to bias . When canopy defoliation takes place during a drought, the NDVI could pick up spectral reflections from the understory which may lead to an underestimate of NDVI change. NDVI could also be biased by pixels covered by cloud. We used MODIS-NDVI product from July and August of 2009 and 2010. As for the reliability of MODIS NDVI data, pixels evaluated as good (class 0) and marginal (class 1) provided reliable data, while data in pixels covered with snow/ice (class 2) or cloudy (class 3) were not reliable. The percentage of pixels with reliability class 0 or 1 was 75.45% and 89.89% for 2009, 70.57% and 88.60% for 2010, respectively, with an average of 81.13%. Only 18.87% pixels were cloudy, which was much lower than the long-term average of over 50% (2001-2014) (Xu et al 2018c). Canopy height is the key data in this research; other spatial data were resampled to match the canopy height data. The canopy height data used here were derived from the LiDAR rather than from the field observations, which was one of the best approaches currently available describing forest vertical structure at regional scales (Simard et al 2011). From known validation at the regional scale, this canopy height data has a good correlation with the field observation tree height in southwestern China reported by Zhang et al (2014a). It is undeniable that this canopy height data also has its own variance and uncertainty, but the canopy heights in our study area (ranged from 15 m to 35 m) were much larger than the regional errors (RMSE = 4.18 m, R 2 = 0.6263). The correlation coefficient (r) between canopy height and elevation is weak (r = 0.349) which modest enough to be both included in the RF model, but it might affect the value of the importance. Gazol et al (2018) suggested that evergreen gymnosperms and deciduous angiosperms have different adaptations and physiological mechanisms to cope with drought, with the former displaying a lower resistance but faster recovery based on tree ring data (TRWi). Although TRWi data were generally more sensitive than NDVI data in terms of accessing resilience to drought (Gazol et al 2018), TRWi data from deciduous broadleaf forest in southwest China is rare and patchy.

Conclusions
In this study, we revealed the characteristics of forest canopy heights related drought resistance based on satellite data and reprocessed climate data. Our results demonstrated that forest resistance to extreme drought depended on canopy height in southwestern China. Taller forests were more vulnerable and susceptible to growth reduction during severe drought than shorter ones. The shorter forests had a greater drought resistance. The frequency and duration of severe drought have been predicted to increase in the near future, more attention should be given to canopy structure for forest management and climate related risk assessments.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files).

Acknowledgments
Thanks to three anonymous reviewers for their constructive improvements to this paper. This study was supported by the National Natural Science Foundation of China (42001348) and Natural Science Foundation of Anhui province of China (2008085QD167). T Zhou was supported by National Natural Science Foundation of China (41571185 and 41621061). X Zhao was supported by National Key Research and Development Program of China (2016YFA0600103). T Zhang was supported by Undergraduate Innovation and Entrepreneurship Training Program of Anhui Normal University (202110370218). H Li was supported by Science and Technology Major Project of Anhui province of China (202003a06020002). We thank Xiaozheng Du for her help on using RF model. We thank China Scholarship Council for its financial support to the first author during 2017-2018, so that all authors have the opportunity to work together on this research. This paper is dedicated to Ruoyang Xu, for she inspired and arrived on the day of its submission.