Spatial Patterns and Drivers of Soil Chemical Properties in a Typical Hickory Plantation

Background: Soil nutrients play critical roles in regulating and improving the sustainable development of economic forests. Consequently, an elucidation of the spatial patterns and drivers of soil nutrients in these forests is fundamental to their management. For this study, we collected 314 composite soils at a 0-30 cm depth from a typical hickory plantation in Lin 'an, Zhejiang Province, China. We determined the concentrations of macronutrients (i.e., soil organic carbon, hydrolyzed nitrogen, available phosphorus, and available potassium) and micronutrients (i.e., iron, manganese, zinc, and copper.) of the soils. We employed random forest analysis to quantify the relative importance of soil-forming factors to predict the soil nutrient concentrations, which could then be extrapolated to the entire hickory region. Results: Random forest models explained 61%–88% of the variations in soil nutrient concentrations. The mean annual temperature and mean annual precipitation were the most important predictor of soil macronutrient and micronutrient concentrations. Moreover, parent material was another key predictor of soil available phosphorus and micronutrient concentrations. Mapping results demonstrated the importance of climate in controlling the spatial distribution of soil nutrient concentrations at ner scales, as well as the effect of parent material, topography, stand structure, and management measures of hickory plantations. Conclusions: Our study highlights the biotic factors, abiotic factors, and management factors control over soil macronutrient and micronutrient concentrations, which have signicant implications for the sustainability of soil nutrients in forest plantations.


Background
Economic forests provide food and timber for human society, and therefore should be managed responsibly, e ciently, and sustainably (Patil 2017). The quantity and quality of economic forests are determined by soil nutrients as well as other factors (e.g., climate) (Littke et al. 2014). Consequently, it is imperative to understand the drivers and spatial patterns of soil nutrients for the e cient and precise management of soil nutrients in these forests (Grimm et al. 2008). For example, soil nutrient mapping may be employed to identify areas where nutrients are de cient and fertilization may be required, and where nutrients are too enriched via environmental overloads, which may have negative impacts (Guan et al. 2017).
There is no consensus on the major driver of soil nutrients for economic forests. The spatial patterns of soil nutrients have been a research focus in the soil and environmental sciences (Liu et al. 2014).
However, due to the high cost of sample collection and analysis, large-scale sampling to obtain the details of the distribution of soil nutrients at regional scales is di cult (Yang et al. 2016). Considerable efforts have been expended in recent years to estimate the spatial variability of soil nutrients and elucidate the causative factors involved across different regions (Wanshnong et al. 2013; Elbasiouny et al. 2014). Owing to the complexity of terrestrial ecosystems, the spatial patterns of soil nutrients vary in Loading [MathJax]/jax/output/CommonHTML/jax.js different regions (Roger et al. 2014;Xin et al. 2016). The spatial distribution of soil nutrients is mediated by the ve state factors of soil formation, namely climate, topography, parent material, organisms, and pedogenic time (Jenny 1941). In addition, stand structure (e.g., stand age and stand density) and management measures (e.g., fertilization and weeding) of economic forests are also closely related to the spatial pattern of soil nutrients (Ronnenberg et al. 2011;Lucas-Borja et al. 2016). These factors vary across different temporal scales and regions, which together affect the spatial variability of soil nutrients (Wang et al. 2009). Consequently, the precise estimation of soil nutrient concentrations across regional scales remains a signi cant challenge ).
Geostatistical methods have been developed to predict the spatial variability of soil nutrients, with the objective of utilizing quanti ed soil properties at a given time and place to predict soil variables at unknown locations (Saito et al. 2005). The promotion of precision forestry and advances in the integration of geostatistical and geographic information systems (GIS) have further evolved the prediction of regional soil nutrients (Lacoste et al. 2014). However, further study is still required to identify the relative importance of different factors and the main controlling factors that affect the spatial variability of soil nutrients. The ensemble approaches of machine learning methods can also be used for the prediction of soil nutrients. Random forest can generate abundant data, which includes information of variable importance and critical variables that control changes in soil nutrients (Liaw et al. 2002).
Random forest has proven to be an effective method for predicting the spatial distribution characteristics and changes in soil organic carbon. This information can be employed to model the soil organic carbon data for each depth interval, to facilitate the comparison of vertical and lateral distribution patterns Hickory (Carya cathayensis Sarg) is an elite subtropical nut and oil tree that is native to China, whose nuts are popular due to their high nutritional value, good taste and unique avor (Wu et al. 2019). Zhejiang Province accounts for more than 70% of the total production of hickory in China, with a total planting area of 86,700 hm 2 . In the main producing area of Lin 'an, hickory accounts for more than 70% of the total income of farmers; thus, it is one of the main economic trees that allows farmers to signi cantly enrich their quality of life. The hickory plantation is restricted by the topographic conditions, with different management methods, and there are also some unmanaged phenomena. To meet the increasing demands for hickory while maximizing its economic bene ts, it is of particular importance to select areas that are highly suitable for its growth (Shen et al. 2016).
There is universally agreed that the potential impact on the spatial patterns of soil nutrients needs to be included in biotic factors (e.g., stand density and stand age), abiotic factors (e.g., climate and topography), and management factors (e.g., fertilization and weeding). Here, we aim to improve our understanding of the variation and the driving factors of the spatial patterns of soil nutrients. Moreover, we hypothesize that the spatial variation of soil nutrients is mainly determined by the climatic factors.
Consequently, it is necessary to fully investigate the spatial patterns of soil nutrients in the main producing areas of hickory so as to master the relationships between impact factors and soil nutrients. Therefore, the objectives of this study were to: 1) identify the controlling factors that drive the spatial Loading [MathJax]/jax/output/CommonHTML/jax.js distribution of soil nutrients. 2) predict and map the spatial distribution of soil nutrients in hickory plantations.

Study area
This study was conducted in Lin'an City (118º ~ 120º E, 29º ~ 31º N), Zhejiang Province, China, which is the central hickory producing area that includes Changhua, Daoshi, Qingliangfeng, and other towns (Chen et al. 2010). This area, is home to a typical subtropical climate with an average annual temperature that

Soil sampling and analysis
For this study, soil samples were collected from 314 sites using the grid method within 1×1 km areas (Fig. 1). Each site was arranged with 10×10 m plots, layout 5 points on the "S"-shaped lines of each plot. Surface soil samples (0 ~ 30 cm) were collected from 5 points and mixed evenly. Approximately 1kg of samples were divided by a quartering method and then transferred to the laboratory for air drying. Meanwhile, data including longitude, latitude, slope, aspect, parent material, stand density and soil type were recorded for each sample site. Moreover, a complete hickory plantation-owner consultation was also carried out to collect the sample sites information such as stand age, fertilization and weeding management measures. Two climate variables (mean annual temperature and precipitation) derived from the WorldClim2 Database at a 1 km spatial resolution (http://worldclim.org/) were used in this study. A distribution map of the sample points in the study area was generated by ArcGIS 10.3.
The soil properties of the different sampling sites were measured based on the standard methods in China, and the soil bulk density was determined using a ring knife method . A hydrometer technique was used to establish the mechanical composition of the soil (Blake 2008), whereas the pH was analyzed using a soil/water ratio of 1:2.5 in an aqueous suspension (Samuels 1976). The soil organic matter (SOM) was determined via wet oxidation using concentrated H 2 SO 4 and K 2 Cr 2 O 7 , and titrating with (NH 4 ) 2 (SO 4 ) 2 ·6H 2 O (Nelson 1996). Based on the assumption that soil organic matter contains 58% carbon, the soil organic carbon concentration was calculated as the soil organic matter concentration × 0.58 (Jackson 1974). Hydrolyzable nitrogen (HN) was hydrolyzed using 0.1 mol L − 1 of NaOH (Wu et al. 2014), whereas soil available phosphorus (AP) was extracted by HCl-NH 4 F and determined by a molybdenum-antimony colorimetric method (Wu et al. 2014). Soil available potassium (AK) was extracted using ammonium acetate and determined by a ame photometric method.

Correlation analysis
Pearson correlation coe cients were calculated to determine the strength of the associations between soil organic carbon, hydrolyzed nitrogen, available potassium, phosphorus, iron, manganese, zinc, and copper. The advantage of the correlation coe cient is that the relationships between variables can be numerically measured and it is directional (Cohen et al. 2009), where 1 represents a positive correlation and − 1 represents a negative correlation. The strengths of the relationships between variables may be quanti ed, and the closer the number is to 0, the weaker the correlation. Correlation analyses were performed in the R software (version 4.0.4), using the direct function cor () calculation and the corrplot package.

Random forest analysis
Data from 314 sampling sites surveyed during our study were analyzed (the frequency distribution of soil nutrient concentrations are depicted in the Additional le (Additional le 1 Fig. S1, Fig. S2). For each soil nutrient data set, we randomly split the data into training and test sets using a 2:1 split. Random forest depends only on three user-de ned parameters: the number of trees (ntree) in the forest, the minimum number of data points in each terminal node (nodesize), and the number of features attempted at each node (mtry). Initially, we tested the combination of ntree, nodesize, and mtry with a training set. More stable results for estimating variable importance were achieved with a higher ntree number (Díaz-Uriarte et al. 2006); thus, we used ntree = 2000, 3000, 5000.
For nodesize we used 3, 5, 7 for regression, which are 3, 5, 7 instances in each terminal node. The default value of mtry in the regression problem is one third of the total number of predictors (p). The predictors we selected included two climate variables: mean annual temperature and mean annual precipitation; two topographical factors: slope and aspect; two stand structure variables: stand age, stand density; two management practices variables: fertilization times, weeding frequency; and parent material. Nevertheless, as the performance of random forest prediction can be sensitive to mtry (Bento et al. 2002;Heung et al. 2014), we applied the mtry values of 1/3p, 2/3p, p. The random forest analysis was then repeated with different parameter combinations for each variable set, and the goodness of t (% var explained) of each combination was compared. We selected the parameter combination with the highest goodness of t. Finally, the data of the training set were predicted by the established model.

Assessment of predictions
The 1/3 test set, namely the out of bag (OOB) sample, primarily uses the common statistical parameters, coe cient of determination (denoted as R 2 oob ), root mean square error (RMSE oob ), and mean absolute error (MAE oob ) to evaluate the random forest model established by the training set. This was calculated by the following formula,

Correlation analysis of soil physicochemical properties in hickory plantation
Correlation analysis is an effective method to reveal the relationship between soil nutrients. The result ( Fig. 2) showed that there was a signi cantly positive correlation between soil organic carbon, available potassium, available phosphorus, hydrolyzed nitrogen and zinc, manganese, iron, which further proved that these nutrients may be affected by similar factors. The correlation coe cient between soil organic carbon and hydrolyzed nitrogen was as high as 0.87. Copper had correlations with the other nutrients with small correlation coe cients, which indicated that copper may have had different driving factors compared with other nutrients in the soil.
Performance and variable importance of random forest models for predicting soil nutrients in a hickory plantation To optimize the performance of random forest predictions in terms of the tting interpretation (% var explained), we used the training set to test the combination of ntree, mtry, and nodesize. The parameter ntree was set to 2000, 3000, 5000, mtry values were 3, 5, 9, and nodesize values were 3, 5, 7. We selected the parameter combination with the highest goodness of t, that is, ntree = 5000, mtry = 3, nodesize = 3 (Additional le 1 Table. S1). In general, the performance of the models was limited. On average, the prediction accuracy was lowest for zinc compared to micronutrient components, namely iron, manganese, copper, and macronutrients ranging from between 0.75 and 0.90 in R 2 oob (Table 1). These results suggested that in the topsoil the spatial distribution patterns of soil nutrients were highly variable Loading [MathJax]/jax/output/CommonHTML/jax.js due to small scale variations in input, redistribution, as well as in the intrinsic random variability of soil nutrients.  (Figs. 3-4). Mean annual temperature and annual precipitation had a strong impact on the prediction of soil organic carbon (Fig. 3a). The level of organic carbon in topsoil is contingent on the inputs of biomass into the soil, which are in uenced by climate. Similar to soil organic carbon, for the prediction of hydrolyzed nitrogen and available potassium, climate was also more crucial than other variables (Fig. 3b, Fig. 3d). Mean annual temperature and parent material both showed high relative importance in the random forest prediction models of available phosphorus (Fig. 3c).
Climate and parent material were also the most critical factors controlling variability of soil iron and copper (Fig. 4a, Fig. 4d). Mean annual precipitation was the dominant variable affecting the spatial distribution of manganese (Fig. 4b), mean annual temperature was the most important factor driving the spatial distribution of zinc (Fig. 4c). Parent material was the second-most important variable that in uenced the manganese and zinc concentration. The role of the parent material was apparent in the spatial distribution of micronutrients. The variables were ranked in the order of climate, parent material, topography, stand structure and management measures. Climate was the most critical predictor for soil nutrients, as it determined their spatial distribution.
Spatial pattern of soil nutrients in a hickory plantation The spatial distribution of soil nutrient concentrations in the hickory plantations mapped by the RF model revealed that all of the soil nutrients had obvious spatial patterns (Figs. 5-6). The concentrations of soil organic carbon and hydrolyzed nitrogen in the soil had similar spatial distribution patterns, with high concentrations primarily located in the west, and obviously low concentrations in the other areas. The concentrations of soil available potassium and available phosphorus in the study area were generally low. The high-value regions of soil available potassium were unevenly distributed exhibited, with the maximum value being 175.10 mg/kg, the minimum value being 118.77 mg/kg. Only a few areas in the northwest of the hickory plantations had the highest soil available phosphorus concentration, which reached up to 15.72 mg/kg.
The soil concentrations of manganese and copper had similar spatial distribution patterns, with high concentrations being located mainly in the northwest and east of the hickory plantation, and obviously low concentration areas in the west and southwest. The iron concentrations in the study area were relatively low. The concentrations of zinc were high, up to 2.44mg/kg, while low-value regions were unevenly distributed.

Discussion
In our study, Random Forest modeling was employed to improve the prediction results, and the optimal settings were selected for each parameter. Through the analysis of each evaluation index, R 2 was as high as 0.88, and the prediction performance of each nutrient model was enhanced. Variable importance revealed different dominating in uencing factors between soil nutrients. Climate variables, namely mean annual temperature and mean annual precipitation were the key factors that drove the spatial changes and concentrations of soil organic carbon, hydrolyzed nitrogen, available potassium. Climate and parent material were the main controlling factor in spatial distribution of soil available phosphorus and micronutrients.
Climate controls the spatial distribution of soil nutrients As hypothesized, climate was the most important predictor of variation in soil nutrients. Correlation analysis revealed that there was a signi cant positive correlation between soil nutrients. The variable importance of soil nutrients was basically the same, which further veri ed that these nutrients may be in uenced by similar factors (Fig. 2). Climate drive the spatial distribution of soil nutrients to a greater degree than do the topography, stand structure and management measures factors. The concentration of soil nutrients is generally negatively associated with temperature and positively associated with the annual mean precipitation Hickory requires appropriate temperatures for growth and su cient water to sustain the physiological processes involved. Soil nutrients in well-developed mountain soils are typically in dynamic equilibrium; however, they are particularly sensitive to climate change (Bowman et al. 2014). When temperatures and water parameters are beyond the range required for optimal growth, soil nutrients can be constrained (Eamus 2003 The proven trend, based on the combination of global data on soils and geology, is that highly weathered soils are more likely derived from acidic or intermediate parent materials, than from calcareous or ma c parent materials (Augusto et al. 2017). Weathered soils derived from acid volcanic rocks were higher in micronutrients than other types of soils. Volcanic rocks are rich in zinc and copper, which can promote the good growth of trees and lay the foundation for obtaining high-quality fruits. Consequently, forest soils from which such parent materials are derived should be given priority when planting hickory. The concentration of soil sand elucidated the spatial distribution of micronutrients better than the soil clay content did, as the sand content is a surrogate of the quartz content, and because the concentration of quartz in rocks is closely related to soil iron and copper (Rahardjo et al. 2004;Bui et al. 2013 In our study, slope also showed a certain relative importance in the spatial prediction of soil available phosphorus. Slope affects runoff, soil moisture concentrations, and the soil erosion rate, which in turn in uences soil nutrients (Johnson et al. 2000;Wang et al. 2009). As the slope angle increases, the precipitation received per unit area, as well as its in ltration decreases due to the greater slope area and higher water ow velocity. Meanwhile, the soil moisture concentration is reduced due to the higher runoff and evaporation area (Florinsky 2016). Consequently, we could not exclude the slope effects even though they were not the most important factor to drive other soil nutrient concentration variabilities in our study. Aspect had low potential in forecasting soil nutrients. One possible explanation for this phenomenon is that (to some extent) climate has an overriding in uence on large scale patterns in ecosystems, including soil carbon cycling, via its control of plant community composition and productivity (Sinoga et al. 2012).
The effects of aspect on soil organic carbon are related to temperature and moisture, and primarily manifested by the sunny aspect being dry, under which the soil organic carbon is decomposed more rapidly, with concentrations lower than that under shady conditions (Zhang et al. 2018).
Stand age in uenced the spatial distribution of soil nutrients more than stand density did. Stand age can play a signi cant role in determining forest oor processes and chemistry. Soil nutrients concentration were mainly affected in three ways: soil nitrogen mineralization, litter decomposition and microbial activity (Welke et al. 2005). The litter input and microbial respiration varied signi cantly under different stand structures. Stand structure can drive soil processes in different ways by altering light, soil temperature and moisture conditions (Hasenauer et al. 2015). However, our study was located in hickory plantation with no obvious difference in stand structure, so the stand age and stand density had little impact on the spatial distribution of soil nutrients.
Our result showed that fertilization and weeding management were not pivotal factors affecting the spatial distribution of soil nutrients, which was inconsistent with management measures being important arti cial means to improve soil nutrient levels (Ronnenberg et al. 2011; Lucas-Borja et al. 2016). The main reason was that the factor used in our analysis was the frequency of fertilization and weeding, not the amount of fertilizer. Because hickory plantation was mostly in mountainous areas, there are many semiarti cial research sites. We investigated 229 sites without fertilization and 102 sites without weeding. In addition, even though some study sites were managed, local farmers had not adopted uniform standards. A variety of fertilizers were used in the production of hickory, including compound fertilizers, nitrogen, phosphorus and potassium fertilizers, and weeding methods contain herbicides or weeding machine.
Therefore, we could only record the frequency of fertilization and weeding. Management measures were vital factors driving the spatial variability of soil nutrients. In random forest modeling and prediction, factors need to be selected reasonably to manifest their effects.
Our study analyzed soil nutrients at a surface depth of 30 cm. The spatial patterns and drivers of deeper subsoils remain unknown and might be important for the production of hickory plantations. However, our results provide a reference for the maintenance and management of soil nutrients in other economic forests. To improve the quantity and quality of economic forest, factors such as climate and parent material should be taken into consideration. Hickory planting sites should have good illumination, Loading [MathJax]/jax/output/CommonHTML/jax.js su cient and uniform rainfall. Soil formed by the weathering of parent rocks that are most suitable for the growth of economic forests should be prioritized.

Conclusions
Through the systematic sampling of soils in a typical hickory region, we quanti ed the relative importance of soil-forming factors to explain regional variations in its physicochemical properties. Macronutrients and micronutrients in hickory plantations all had obvious spatial patterns. The mean annual temperature and mean annual precipitation were found to be the most signi cant factors for elucidating the spatial variations in soil nutrients. Climate and parent material were the critical factor involved in controlling the spatial variations of soil available phosphorus and soil micronutrients. Slope was also important for explaining the spatial variations in soil available phosphorus and iron. Aspect, stand structure and management measures were less important in the prediction of spatial variations for both soil macronutrients and micronutrients but should not be ignored. An improved understanding of the spatial variations and drivers of soil nutrients in plantations will aid in the development of effective strategies for soil nutrient sustainability.

Declarations
Ethics approval and consent to participate Not applicable.

Consent for publication
Not applicable.

Availability of data and materials
The data that support the ndings of this study are available from the corresponding author upon reasonable request.  Mapping spatial distribution maps of macronutrients in soils. SOC indicates soil organic carbon, HN indicates soil hydrolyzed nitrogen, AP indicates soil available phosphorus, AK indicates soil available potassium.