Hierarchical determinants of winter wheat abandonment in the North China Plain: A case study of Xingzhuangzi village in Hebei Province

Highlights

•	Plot level and household level were both crucial in explaining farmers’ abandonment of winter wheat.
•	Winter wheat was more likely to be abandoned on plots with lower land quality and larger plot areas.
•	Winter wheat was less likely to be abandoned on plots with stable water availability.
•	Household non-agricultural income ratio played a positive role on farmers’ abandonment of winter wheat.
•	Household demographic characteristics also contributed significantly to farmers’ abandonment of winter wheat.

Abstract

Since the late 1990s, the North China Plain (NCP) has undergone large-scale shrinkage of area sown to winter wheat, accompanying with traditional double cropping system being replaced by spring corns. However, studies on the underlying determinants are rarely found. The goal of this paper is to detect the hierarchical determinants on farmers’ cropping system decisions. A case study was carried out in Xingzhuangzi village of Hebei province, and multi-level statistic models were constructed using household survey data. Results show that plot level and household level were both crucial in explaining farmers’ land use decisions: winter wheat was more likely to be abandoned on plots with lower land quality, unable to be irrigated, and with larger plot areas; at household level, both the non-agricultural income ratio and the land fragmentation played positive roles on farmers’ abandonment of winter wheat while the role of household agricultural labor availability was negative. There was also a nonlinear relationship between average age of households’ agricultural laborers and their cropping system decisions, and middle-aged farmers had a lower probability to abandon winter wheat. Overall, this paper provides empirical identification on hierarchical determinants of agricultural land use change in the NCP, and encourages policies aiming at adjustment of cropping systems, integration management of both surface and groundwater, and promotion of land transfer, in order to achieve the twin goals of ecological conservation and food security in water-scarce areas.

Keywords:

• Agricultural land use change
• Multi-level modelling
• Water availability
• Household agricultural labor availability
• Water-scarce areas

1 Introduction

Agricultural land use changes are closely related to the social-economic development and have important consequences as agricultural areas provide a wide range of goods and services [1,2]. Understanding agricultural land use change processes and their drivers is crucial for improvement of future planning strategies and assessment of the influence of land related policies [2,3].

China is a nation with strong rural roots [4]. Since the economic reforms and open-door policy were initiated in 1978, China has experienced fast economic development, and extensive rural labor has been migrating into the secondary and tertiary sectors [5,6]. Meanwhile, the agricultural land has undergone tremendous changes, varying in usage type and in intensity, mainly in the way of cultivated land conversion in peri-urban areas, land abandonment in sloping areas and cropping system changes in plain areas [7,8]. More specifically, changes of double cropping rice to single cropping rice have been specified in plain areas of southern China [9], while the North China Plain (NCP) have faced winter wheat abandonment and its accompanying cropping system changes [10].

The NCP is one of the major food production areas in China [11]. Winter wheat – summer maize double cropping has been the dominant cropping system in the NCP since the 1970s [12]. However, recent studies have found that large, spatially-continuous areas in the NCP which were previously winter wheat – summer maize double cropping system are being replaced by the single cropping system of spring corn; this has been termed “the spring corn planting belt phenomenon” [10,13,14]. Since winter wheat is a major cereal crop, the cropping system changes must have an impact on food security. The NCP is also amongst the global “hotspots” in terms of water scarcity and groundwater depletion [15], and winter wheat plays a key role in the development of water scarcity and the over-exploitation of groundwater, as more than 70% of irrigation water, i.e. the vast majority of total water use in the NCP, is consumed by winter wheat [16,17]. Therefore, the abandonment of winter wheat concurrently affect water balance of the NCP. In this context, identification of its drivers is critical to achieve the twin goals of food production and water conservation in the NCP.

Previous studies concerning determinants of agricultural land use changes in China have focused mainly on the impact of urbanization on cultivated land conversion [8,18,19]. There are also several studies relating to driving factors of sloping land abandonment in western China and rice cropping system changes in southern China [20–22^[20][21][22]]. Most of the results have indicated that the above mentioned rural labor out-migration is a critical element in addressing the agricultural land use changes, accompanying with other elements, including physical contexts, social-economic characteristics, policy-relevant indicators and so on [8,22]. However, to our knowledge there has been no study that has analyzed the determinants of winter wheat abandonment in the NCP [23].

In addition, household survey data are widely applied in empirical studies, since agricultural land use changes often result from the decisions of individual farming households, and conventional regression techniques, notably the multiple linear regression model, the seemingly unrelated regression model, the Probit model and the Tobit model, are popular approaches to explore datasets and to test hypothesis between land use change variables and potential determinants [5,20,22,24]. However, farmers’ land use decisions are influenced by diverse interactive factors at multiple levels, for example physical contexts at plot level, demographic and social-economic characteristics at household level and policy-relevant indicators at macro levels, and conventional regression models are not optimal in resolving such nested relationships [21,25].

Multi-level modelling was introduced for the analysis of hierarchically structured data, which is a statistically sound methodology with regression models that explicitly takes variability at different levels into account [26,27]. Multi-level models have mainly been used in the social and medical sciences, and are becoming popular in geographical applications [28,29]. Multi-level models also hold great promise in understanding agricultural land use decisions with nested hierarchical structures. However, to date, there have been few empirical examples that applied this type of model to such analysis [21,25,30].

Considering the above shortcomings, this paper aims to unravel the hierarchical determinants of winter wheat abandonment in the NCP through a case study of Xingzhuangzi village in the Hebei province. For this, a multi-level statistical model is constructed to examine the relative importance of determinants on cropping system decisions at plot and household levels, using first-hand data collected by means of household surveys. As a case study, this paper is useful in reflecting a broad picture of cropping system decisions in the NCP, and also in giving support on the development and implementation of policies on agricultural land use and water management, in order to alleviate water stress while maintaining food supply in water-scarce areas of China.

2 Materials and methods

2.1 Study area

Xingzhuangzi village is located in Cangzhou of Hebei province (Fig. 1). It is a large village with about 700 households and 2000 people, and the area of cultivated land is more than 400 ha. This village enjoys a temperate monsoon climate, with an annual average temperature of 13 °C, and an annual average precipitation of 590 mm. Winter wheat – summer maize double cropping system used to be the dominant cropping system. However, single cropping has spread across this village since late 1990s, and spring maize was the most popular choice [13]. Therefore, the major crops for this village are winter wheat, summer maize and spring maize, accounting for nearly 90% of total area of cultivated land. In addition, there are also small areas of cotton, sweet potato, and sorghum.

Fig. 1 Location of the study area.

2.2 Data collection

A two-stage survey was conducted between April and May in 2015, in order to collect the data necessary for multi-level modelling of farmers’ cropping system decisions. The first stage was a preliminary survey using interviews with dozens of county or village officers, field tours and Landsat TM images in Cangzhou. After this survey, Xingzhuangzi village was selected for detailed study, since it has a typical phenomenon of winter wheat abandonment. The second stage was the household interviews in Xingzhuangzi village. Questionnaires were completed through semi-structured one-on-one interviews with the household heads or the household members responsible for the farm. The households were selected randomly in the village, and those with qualified members absent were skipped. In all, 100 households were interviewed, covering 472 land plots with a total area of 90.23 ha.

The questionnaire obtained abundant information on households and the plots with which they were associated. At plot level, we recorded physical information for each land plot, including land quality, irrigation water source, plot area, distance from the land plot to the residence of the household and also an input-output table for crops. At household level, information was collected mainly on demographic characteristics and household income, including the age, gender, education level, marital status, occupation and income for each household member. Additionally, we recorded farmers’ perceived determinants for winter wheat abandonment on single-cropping land plots. In this paper, only a subset of the data were used.

2.3 Multi-level model specification

We followed the common approach suggested in the literature to build a series of multi-level models, testing at each step the tenability of the hypothesized level-based variation [31,37]. First, we started with Model 1 incorporating only group effects. A logistic model is appropriate, as there are two main cropping systems in the study area, i.e., the spring maize single cropping system (single cropping system) and the winter wheat-summer maize double cropping system (double cropping system). Here, 1 was used to express the single cropping system, and 0 was used to express the double cropping system; p was the probability of the occurrence of a “single cropping system” event.(1) $\text{Log}\left(\frac{P_{ij}}{1-P_{ij}}\right)={\gamma }_0+u_j+{\epsilon }_{ij}$(1)

In Eq. (1(1) $\text{Log}\left(\frac{P_{ij}}{1-P_{ij}}\right)={\gamma }_0+u_j+{\epsilon }_{ij}$(1) ), γ₀ is the fixed intercept to be estimated. u_j is the random intercept that varies randomly between groups, i.e. households in this paper, and is assumed to be normally distributed with a zero mean and a variance of τ² [27]. ε_ij is the deviation between individual i, and is also normally distributed with a zero mean and a variance of σ². For all multi-level models in this paper, the following notations are applied: i indexes land plots and j indexes households.

Model 1 is a two-level random intercept model, and is also called the “empty model”, since it contains no explanatory variable [25]. This model decomposes the variance of the dependent variable into two parts: one caused by the plot level and the other caused by the household level. It is equivalent to a one-way ANOVA with random effects, and is used as a base model to test whether higher-level variance is significant [31].

Next, we sequentially added two sets of explanatory variables to Model 1, developing Model 2 and Model 3, in order to estimate the effects of the indicators at plot and household levels on farmers’ cropping system decisions. When including all explanatory variables, a random-intercept two-level model takes the following form:(2) $\text{Log}\left(\frac{P_{ij}}{1-P_{ij}}\right)={\gamma }_0+{\displaystyle {\displaystyle \sum }_{m=1}^M}{\alpha }_m{X}_{mij}+{\displaystyle {\displaystyle \sum }_{n=1}^N}{\delta }_n{Z}_{nj}+u_j+{\epsilon }_{ij}$(2)

In Eq. (2(2) $\text{Log}\left(\frac{P_{ij}}{1-P_{ij}}\right)={\gamma }_0+{\displaystyle {\displaystyle \sum }_{m=1}^M}{\alpha }_m{X}_{mij}+{\displaystyle {\displaystyle \sum }_{n=1}^N}{\delta }_n{Z}_{nj}+u_j+{\epsilon }_{ij}$(2) ), X_mij consists of m variables at plot level, and Z_nj contains n variables at household level; and α_m and δ_n are the corresponding regression coefficients.

Eq. (2(2) $\text{Log}\left(\frac{P_{ij}}{1-P_{ij}}\right)={\gamma }_0+{\displaystyle {\displaystyle \sum }_{m=1}^M}{\alpha }_m{X}_{mij}+{\displaystyle {\displaystyle \sum }_{n=1}^N}{\delta }_n{Z}_{nj}+u_j+{\epsilon }_{ij}$(2) ) contains random intercept effect at household level (u_j), but no random slope effect. This is because the variances of random slopes were not significant when estimated. Considering that the data structure of this paper is characterized by the observation of 472 land plots nested within 100 households, the number of observations (land plots) per household is small and may hamper the estimation of random slopes. This finding is similar with previous studies which also identified insignificant random slopes or models that did not converge [21,25,29]. Since no argument suggested that the relationship between the explanatory variables and dependent variable are different when included random slopes, we employed multi-level models without random slopes in this paper.

The analyses were performed using the multi-level statistical package HLM version 7 [32], and three multi-level models were all estimated applying a restricted maximum likelihood (RML) algorithm [27]. The intra-class correlation coefficient (ρ) were calculated to indicate the proportion of variance explained by the household level [33], using the following equation:(3) ${\rho }_2=\frac{\text{VAR}\left(u_j\right)}{\text{VAR}\left(u_j\right)+\text{VAR}\left({\epsilon }_{ij}\right)}$(3)

In addition, the relative operating characteristic (ROC) was used in order to assess the goodness-of-fit of the models [21]. Models will show better fitting effects with the ROC values increasing from 0.5 (completely random) to 1.0 (perfect discrimination).

2.4 Potential explanatory variables

In this paper, potential explanatory variables mainly cover the characteristics at plot level and household level (Table 1).

Table 1 Description of the variables used in the multi-level model.

Variable Name	Description	Mean	SD	Min	Max
Dependent variable (plot level, n = 472)
Cropping system	1 if spring maize single cropping system, 0 if winter wheat − summer maize double cropping system	0.49	0.50	0.00	1.00
Explanatory variables
Level 1 (plot level, n = 472)
Land quality	1 good land; 2 relatively good land; 3 relatively poor land; 4 poor land	2.22	1.04	1.00	4.00
Irrigation water source
Ground & surface water	1 if the land plot can be irrigated by both groundwater and surface water; 0 otherwise	0.14	0.34	0.00	1.00
Groundwater only	1 if the land plot can be irrigated by only groundwater; 0 otherwise	0.34	0.47	0.00	1.00
Surface water only	1 if the land plot can be irrigated by only surface water; 0 otherwise	0.39	0.49	0.00	1.00
(No irrigation)	Reference group
Plot area	Area of the land plot (ha)	0.17	0.09	0.01	0.53
Distance to residence	Distance from the land plot to the residence of the household (km)	1.00	0.85	0.02	7.50
Level 2 (household level, n = 100)
Non-agricultural income ratio	Ratio of non-agricultural income to total income in the household	0.57	0.24	0.00	0.88
Household agricultural laborer availability	Number of agricultural labor per unit land area for the household (laborer/ha)	2.70	1.55	0.60	8.33
Average age	Average age of agricultural laborers in the household (year)	54.35	9.14	35.00	74.00
Male age	Average age of male laborer stayed in agriculture (year)	55.04	12.36	24.00	81.00
Female age	Average age of female laborer stayed in agriculture (year)	50.55	11.27	25.00	72.00
Average age squared	Squared average age of agricultural laborers in the household (square year)	3036.33	991.05	1225.00	5476.00
Average education level	Average education level of agricultural laborers in the household: 1 illiteracy; 2 primary school; 3 middle school; 4 high school; 5 college	2.24	1.79	1.00	5.00
Ratio of male laborer	Ratio of male among agricultural laborers in the household	0.30	0.19	0.00	1.00
Land fragmentation	Number of plots per unit land area for the household (plot/ha)	6.09	1.95	2.00	15.00

The plot level variables include land quality, irrigation water source, plot area and distance to residence. Land quality was categorized into four classes: good, relatively good, relatively poor and poor land, based on output capacity. Considering the irrigation water source, there are four types of land plots, i.e., those can be irrigated by both ground- and surface water, those by only groundwater, those by only surface water and those cannot be irrigated, with a probable declining water availability, as surface water is not stable compared to groundwater in the NCP [34,35]. As shown in Table 1, irrigation water source contains three dummy variables with no irrigation as a reference group, aiming to specify the influence degrees of different irrigation water sources on farmers’ cropping system decision. Plot area was initially recorded using the unit of mu, which was then converted to the unit of ha (1 ha = 15 mu). The distance to residence was the actual walking distance between the land plot and its owner’s residence.

As mentioned above, rural labor out-migration usually act as a crucial factor on farmers’ land use decisions, and its influences are mainly manifested in two aspects [5,36]. The first one is through the lost-labor effect. Due to rural labor out-migration, households with insufficient agricultural laborer are more likely to choose extensive land use. The second one is through the income effect. Off-farm employment, accompanying with rural labor out-migration, is expected to increase household income, and thereby influence farmers’ land use decision. In this context, two variables relating to rural labor out-migration, i.e., non-agricultural income ratio and agricultural labor availability, were analyzed at household level. In addition, four indicators of the household agricultural labor resources were employed, including the average age, the average age squared, the average education level, and the ratio of male laborer. Here, we used both average age and average age squared, because average age may show a non-linear relationship to agricultural land use changes [21]. We also incorporated land fragmentation at household level, defined as the number of plots per unit land area. As higher fragmentation means more inconvenience for land farming, we expected land fragmentation to positively influence farmers’ preference for single cropping system.

3 Results and discussion

3.1 Background of the sample land plots and households

In the study area, a total of 100 households and 472 land plots were investigated. The land use share was 49.15% (232) of the land plots farming with single cropping system, and 50.85% (240) with double cropping system (Table 2). The land plots are often small and fragmented, and the average area is only 0.17 ha (Table 1). The proportions of plots with good, relatively good, relatively poor and poor land quality are 32.20%, 26.69, 27.54 and 13.56, respectively, and 13.77% of the plots cannot be irrigated (Table 2). The value of distance to residence for each land plot ranges from 0.02 km to 7.50 km with an average of 1 km (Table 1).

Table 2 Distribution of the land plots by cropping system, land quality and irrigation water source.

		No. of plots	% of total plots
Cropping system	Single cropping system	232	49.15
	Double cropping system	240	50.85
Land quality	Good	152	32.20
	Relatively good	126	26.69
	Relatively poor	130	27.54
	Poor	64	13.56
Irrigation water source	Ground & surface water	65	13.77
	Groundwater only	159	33.68
	Surface water only	183	38.77
	No irrigation	65	13.77

At household level, the average cultivated land area owned by each household is 5.32 plots and 0.92 ha (Table 3). In terms of household members, 265 of the total 498 individuals are male and 233 are female. The number of laborer (aging from 16 to 65) is 338, and 55.03% of the laborers are mainly engaged in non-farm activities (Table 3). For household members engaged in farm work, the average age is 54.35, and the average education level is 2.24, which is between 2 (primary school) and 3(middle school) (Table 1).

Table 3 Characteristics of the households and their members.

	Value
Average number of land plots	5.32
Average cultivated land area/ha	0.92
Number of males	265
Number of females	233
Number of labors	338
Number of labors mainly engaged in non-farm activities	186

3.2 Farmers’ perceived determinants

Fig. 2 presents farmers’ perceived determinants linked to winter wheat abandonment on single-cropping land plots.

Fig. 2 Farmers’ perceived determinants for winter wheat abandonment on single-cropping plots.

Note: Winter wheat abandonment may have more than one reason. For example, a farmer may abandon winter wheat on one plot because of both small plot and labor shortage. Therefore, the total number of land plots (271) in Fig. 2 are larger than that of the total number of single-cropping plots (232).

From the viewpoint of farmers, the most important determinant for abandoning winter wheat is the land plots’ low water availability. Besides, economic factors, poor land quality and shortage of labor play important roles in farmers’ winter wheat abandonment decision. Here, economic factors include high cost of fertilizer, high cost of pesticide and low price of winter wheat. In addition, small plot was regarded as the critical determinants in two land plots. There were also other causes including rough land plot, “peer effects” (one abandons winter wheat following the others’ abandoning), etc. It is worth mentioning that no household chose long distance as a determinant, implying that the influence of distance to residence was likely to be insignificant on farmers’ cropping system decision.

3.3 Econometric results of the multi-level models

Table 4 displays the estimated results of the multi-level models, with different sets of explanatory variables incorporated.

Table 4 Multi-level models for cropping system decisions.

Variables	Model 1	Model 2	Model 3	Model 4
Fixed effect
Level 1 (plot level)
Intercept	−0.125	−0.114	8.366	8.511
Land quality		0.655^***	0.731^***	0.800^***
Irrigation water source
Ground & surface water		−2.727^***	−3.195^***
Groundwater only		−2.551^***	−3.019^***
Surface water only		−2.212^***	−2.666^***
(No irrigation)
Water availability				0.718^***
Plot area		−4.050^***	−4.359^***	−4.015^***
Distance to residence		−0.033	−0.051	−0.093
Level 2 (household level)
Non-agricultural income ratio			1.138^*	1.018^*
Household agricultural labor availability			−0.447^***	−0.421^***
Average age			−0.348^*	−0.354^*
Average age squared			0.003^*	0.003^*
Average education level			−0.134	−0.110
Ratio of male laborer			−0.965	−0.813
Land fragmentation			0.280^***	0.259^***
Random effects
Level 2
VAR(R)	0.627^***	2.475^***	2.127^***	1.524^***
ρ	0.408	0.775	0.742	0.654
ROC	0.816	0.930	0.927	0.910

* p < 0.1.

*** p < 0.01.

Results of Model 1 showed that the random effect at household level was significant, suggesting a significant difference in farmers’ cropping system decisions between households. The intra-class correlation coefficient ρ was 0.408. Therefore, 40.8% of the total variance should be attributed to the variation within households, reflecting a considerable clustering of the occurrence of single cropping, i.e. abandonment of winter wheat, at household level.

Model 2 considered fixed effects at plot level, and incorporated explanatory variables including land quality, three dummy variables for irrigation water source, plot area and distance to residence. All variables, except distance to residence, contributed significantly to the probability of farmers’ choice on single cropping system. The signs of their coefficients were consistent and the changes in their magnitudes were minor, even when we added more explanatory variables in Model 3 (Table 4).

In addition to variables at plot level, variables at household level were also incorporated in Model 3. Except for average education level and ratio of male laborer, the other variables all showed significant contributions.

In the above three models, the random effects at household level were all significant. As expected, the ROC value increased from Model 1 to Model 3. This also confirmed that the explanatory variables at plot and household levels did explain part of variances in farmers’ cropping system decisions. Next, the results of Model 3 will be used to analyze the specific effects of each variable on farmers’ cropping system decisions, at plot level and household level, respectively.

3.3.1 Land quality

The coefficients of land quality was significant and positive, representing that land plots with poorer quality were more likely to suffer winter wheat abandonment and to adopt spring maize single cropping system. This is in accord with the empirical study by Zhang et al. [21], who pointed out that the cultivation cost of land with poor quality usually cannot be covered by the profit from its yield and therefore the land plot is prone to extensive utilization.

3.3.2 Irrigation water source

Taking no irrigation as the reference group, Model 3 estimated that three types of irrigation conditions, i.e. irrigated by both ground- and surface water, only groundwater and only surface water, all showed significant negative effects on farmers’ abandonment of winter wheat, and the absolute values reduces from 3.195 to 2.666. As mentioned above, the availability of surface water is much lower than that of groundwater in the NCP [34,35]. Considering the land plots that can be irrigated, contrast to those with high water availability sourced from groundwater, those with low water availability sourced from only surface water were more likely to suffer abandonment of winter wheat. Meanwhile, land plots with no irrigation water source suffered largest share of winter wheat abandonment, which coincided with the well-known local proverb: No irrigation, no winter wheat, since irrigation water is essential for the growth of winter wheat in the NCP [10].

3.3.3 Plot area

The estimation results of Model 3 showed that the influence of plot area was significant, and the coefficient was negative. Therefore, the larger the area of a plot is, the less likely it will suffer abandonment of winter wheat, and the more likely it will adopt double cropping system. This confirms the empirical analyses by [37], who also revealed that plot area played significant negative role in extensification land utilization in Shandong province of the NCP.

3.3.4 Distance to residence

The distance to residence did not play significant effect on farmers’ cropping system decisions. According to empirical studies in mountainous areas, the influence of distance to residence on farming was significant and remote land plots with more labor input and less land rent were more likely to be abandoned [21,22], which is different with the finding in this paper. This can be explained by that in plain areas such as the NCP, the influence of distance to residence on labor input and land rent is minor and farmers did not take the distance to residence into consideration when they decided whether or not to grow winter wheat on the plot. Results of farmers’ perceived determinants also confirmed the insignificant influence of this indicator (Fig. 2).

3.3.5 Non-agricultural income ratio

The variable non-agricultural income ratio influenced positively on farmers’ choice of single cropping system. Larger ratios of non-agricultural income to total household income were correlated with less probabilities of farmers to allocate their land for double cropping. This finding reveals that with the out-migration of rural labor, non-agricultural income becomes an important component in total household income, which significantly discouraged farmers’ initiative to grow more winter wheat. Farmers prefer buying wheat products from the market using the non-agricultural income. This is also confirmed by the empirical study by Kokoye et al. [24].

3.3.6 Household agricultural labor availability

According to [21,38], the number of agricultural laborers within the household indicates the household agricultural labor availability, which should intensify the use of agricultural land. In this paper, household agricultural labor availability did play a significant and negative role in farmers’ choice of single cropping. That is, with the out-migration of rural labor, the more labor staying for agricultural, the less probability will their land been abandoned for winter wheat.

3.3.7 Average age of agricultural laborers and its quadratic form

In Model 3, the estimated coefficients for average age and average age squared were both significant and showed opposite effects, with a positive coefficient for the quadratic form but a negative coefficient for the root form. This implies that households with higher average ages of agricultural laborers tend to experience less abandonments of winter wheat but the changing rate declines as the average age increases. Therefore, middle-aged agricultural laborers showed low probabilities of abandoning winter wheat on their plots, while too young or too old laborers showed larger probabilities. This corroborated the finding by [21], who revealed that households with agricultural laborers that are too young or too old would adopt extensive land use because of lack of cultivation experience or limited physical strength.

3.3.8 Average education level of agricultural laborers

The coefficient of the average education level was not statistically significant, implying that there was no obvious difference in cropping system decisions among agricultural laborers with diverse education levels. This also confirms the empirical findings by Kokoye et al. [24] and Su et al. [30].

3.3.9 Ratio of male laborers among agricultural laborers

The ratio of male laborers also showed an insignificant influence on the occurrence of winter wheat abandonment. As mentioned in Section 3.1, the average ages of male and female agricultural laborers were 55.04 and 50.25 years, respectively, and the remaining male agricultural laborers were much elder than the female ones, due to the out-migration of larger proportion of young, strong male laborers [8,39]. In this context, the difference in cropping system decisions were minor between male- and female agricultural laborers.

3.3.10 Land fragmentation

As expected, Model 3 showed a significant and positive coefficient for land fragmentation, households with higher land fragmentation were more likely to adopt single cropping system. This coincided with the results of empirical studies by Niroula and Thapa [40] and Deininger et al. [41], revealing that land fragmentation was detrimental to agricultural farming and economic gain, thereby triggered lower production inefficiencies and discouraged farmers from intensive land use.

4 Policy implications

Based on the above results, plots with poor land quality were more likely to suffer winter wheat abandonment. On the one hand, the spatial variation of land quality was a result of sediment changes caused by swing and deformation of the riverbeds during historical period; on the other hand, farmers’ cultivation behaviors also impacted land quality [42,43]. According to the investigation results of soil fertility of cultivated land by China’s Ministry of Agriculture in 2014, land degradation was found in the NCP due to long-time over-exploitation of land resource and excessive use of fertilizers [44]. In addition to soil amelioration, land formation retirement on plots with poorer land quality, by leaving land fallow during the sown period of winter wheat winter, has become a popular choice to reserve land fertility in the NCP. However, in order to ensure food supply of the NCP, triple cropping system in two years, instead of single cropping system, may be a better solution, and is suggested to be promoted actively by local government.

From the point of irrigation condition, ground- and surface-water irrigation are both beneficial for winter wheat adoption. Considering the severity of groundwater over-exploitation problems, expansion of areas irrigated by surface water should be a tradeoff between conservation of groundwater and guarantee of food security in the NCP. However, the supply of surface water is erratic [45]. In this context, surface water management strategies should be formulated, and measures for surface water utilization, for example rainwater harvesting and biotechnological advances [46], are strongly recommended. Close attention should also be paid to the improvement of groundwater management; collecting groundwater resource charges, the use of brackish water, and sprinkle irrigation are all possible strategies to reverse groundwater over-exploitation and to promote groundwater storage [47,48].

Taking into consideration of both plot area, land fragmentation and variables of household labor resources, double cropping system is more likely to be adopted in households with larger area of plots, low land fragmentation and more middle-aged agricultural laborers. Therefore, land transfer, from households with agricultural labor shortage to those with ample agricultural laborers, is conducive to plot mosaic and full utilization of labor force, and can therefore alleviate abandonment of winter wheat. This is consistent with the Comments on guidance of the rural land transfer management and development of agricultural moderate-scale operation, issued by the central government, which also aims to raise labor productivity and to ensure national food security [49]. However, the land transfer ratio in the study area is only 2%, much smaller than the average ratio in China (30.4%) [50]. Policies, concerning protection of land ownership, improvement of land market, and land registration system, are suggested, in order to ensure easy access to land rent market and to increase longer-term land transfers.

5 Conclusion

In this paper, we applied multi-level models to reveal the hierarchical determinants at plot and household levels that affect farmers’ decisions on whether or not to grow winter wheat. The results revealed that plot level and household level were both crucial in explaining farmers’ land use decisions; hierarchical determinants concerning plot characteristics, including land quality, water availability and plot area, showed significant influence on farmers’ choice of double or single cropping system; at household level, both the non-agricultural income ratio and agricultural labor availability of the household contributed significantly to farmers’ cropping system decisions. We also found that the average age of agricultural laborers showed a nonlinear relationship, and middle-aged agricultural laborers were more likely to adopt double cropping system. Additionally, land fragmentation played a positive role in farmers’ abandonment of winter wheat.

This paper presented that the multi-level modelling methodology had a good performance in identifying hierarchical determinants on farmers’ cropping system decisions. The methodology can also be applied into other cases concerning hierarchically structured activities underlying regional agricultural land use changes. In addition, we recommended policies on adjustment of cropping systems, integration management of water resources and promotion of land transfer in the NCP, based on the main findings of this paper. The recommended policies are also appropriate in similar water-scarce areas, for the twin goals of ecological conservation and food security.

Concerning the limitations, this paper focused mainly on the determinants at plot level and household level, since we collected household survey data within a typical village. Further analysis with additional samples is still need, for the purpose of revealing comprehensively the hierarchical determinants of winter wheat abandonment in the NCP.

Acknowledgements

This work was supported by the NSFC-IIASA International Cooperation and Exchange Program (41161140352) and the Natural Science Foundation of China (41271119). We thank Yuluan Zhao (Guizhou Normal University) for his suggestions during the study. We also thank the editor and the anonymous reviewer for their insightful and constructive comments on earlier versions of the manuscript.