From statistics to grids:A two-level model to simulate crop pattern dynamics

XIA Tian ,WU Wen-bin ,ZHOU Qing-bo ,Peter H.VERBURG ,YANG Peng ,HU QiongYE Li-ming,ZHU Xiao-juan

1 Key Laboratory for Geographical Process Analysis &Simulation,Hubei Province/College of Urban &Environmental Science,Central China Normal University,Wuhan 430079,P.R.China

2 Key Laboratory of Agricultural Remote Sensing,Ministry of Agriculture and Rural Affairs/Institute of Agricultural Resources and Regional Planning,Chinese Academy of Agricultural Sciences,Beijing 100081,P.R.China

3 Agricultural Information Institute,Chinese Academy of Agricultural Sciences,Beijing 100081,P.R.China

4 Institute for Environmental Studies,VU University Amsterdam,Amsterdam 1085,The Netherlands

5 Department of Geology,Ghent University,Ghent 9000,Belgium

6 Commercial and Economic Law School,China University of Political Science and Law,Beijing 100088,P.R.China

Abstract Crop planting patterns are an important component of agricultural land systems. These patterns have been significantly changed due to the combined impacts of climatic changes and socioeconomic developments. However,the extent of these changes and their possible impacts on the environment,terrestrial landscapes and rural livelihoods are largely unknown due to the lack of spatially explicit datasets including crop planting patterns. To fill this gap,this study proposes a new method for spatializing statistical data to generate multitemporal crop planting pattern datasets. This method features a two-level model that combines a land-use simulation and a crop pattern simulation. The output of the first level is the spatial distribution of the cropland,which is then used as the input for the second level,which allocates crop censuses to individual gridded cells according to certain rules. The method was tested using data from 2000 to 2019 from Heilongjiang Province,China,and was validated using remote sensing images. The results show that this method has high accuracy for crop area spatialization. Spatial crop pattern datasets over a given time period can be important supplementary information for remote sensing and thus support a wide range of application in agricultural land systems.

Keywords:crop planting pattern,spatialization,simulation,spatiotemporal change,remote sensing

1.Introduction

Agricultural land systems are important terrestrial components of the earth system and encompass all the activities and processes related to hu man cropland use. Agricultural land provides the majority of the global food supply;more than 90% of all food calories and approximately 80% of all food protein and fats are directly or indirectly derived from agricultural land systems (Foleyet al.2005;Yeet al.2013;Wuet al.2014). With the increase in the global population and climate change,food security has become a major strategic issue for national economic development and social stability (Yeet al.2016). Food security research focuses on changes in crop planting patterns. The mechanism of changes in crop planting structure helps to explore the feedback of crop patterns to climate change and policy impacts.The research results will improve agricultural policies and ensure regional food security. It is thus of great importance to understand the spatiotemporal dynamics of agricultural land systems and to develop effective management strategies to increase future food production while simultaneously protecting the agroecosystem.

The spatial pattern of crops can reflect many service functions within the agricultural land system,such as food security,cropland carbon sequestration and bioenergy production (Yeet al.2010). In addition,it also reflects the human utilization of agricultural production resources in space. Through research of the processes involving the changes in crop spatial patterns,the spatial evolution of agricultural crops can be further explored,which provides a basis for the adjustment and optimization of the cropping structure. There has been much research investigating the dynamic changes in agricultural land systems over time and space. Satellite images provide an important tool to map the spatio-temporal dynamics of crop patterns. Yet,there are still challenges in identification of crop types using remote sensing and long time-series crop maps are still not available due to the limited time coverage (Pilehforooshhaet al.2014;Melo de Oliveira Santoset al.2019). At this stage,most of the agricultural data are statistical data,which are collected through statistical surveys and regional reporting. These data can reflect the regional aggregates of the planting area,but are insufficient to support production decisions at sub-regional scales in situations such as precision farming. In other words,the statistics data can provide crop area information in a region,however,these data are not suitable for the adjustment of cropping structure. The lack of high precision and longterm crop pattern data is a stumbling block to agricultural research in,climate change and food security,where crop planting structure adjustment cannot be spatially analyzed in line with the crop planting patterns. These studies can only be carried out using correlation analysis at the regional scale,while impact studies are difficult to carry out at the pixel scale. Spatially explicit modeling techniques have emerged as an alternative to monitoring approaches because they are capable of representing landscape allocation and possible developments (Xiaet al.2014). Agricultural land use models are becoming an important tool both for conceptualizing and testing scientific hypotheses regarding the role of different land use change drivers and for exploring future development scenarios(Verburget al.2006;Waiyasusriet al.2016;Xiaet al.2016;Zhanget al.2016). These models have also played a major role within land system science for structured analyses of the complex interactions within the land system(Rounsevellet al.2012).

A wide variety of agricultural land use models have been developed to address research questions related to various processes at a range of scales (Heistermannet al.2006;Bettset al.2015;Mirkatouliet al.2015;Qureshiet al.2018;Nielsenet al.2019). Increases in computational power have extended the capabilities of geographic information systems (GIS) to analyze and model spatial dynamics (Voinovet al.1999). Several agricultural land system models have integrated modules that simulate farmer management practices and decisionmaking processes (Noszczyk 2019). For instance,Temme and Verburg (2011) mapped and modeled the agricultural land-use intensity estimated from nitrogen inputs across the European Union. A multiscale modeling approach exploring the spatiotemporal dynamics of European livestock distribution was proposed by Neumannet al.(2011). However,these models have mainly been designed to simulate the changes and transitions between generic land-cover types,such as cropland,forest,grassland and built-up areas,while changes in crop planting patterns within croplands are generally ignored (Andersonet al.2015;Eitelberget al.2015;Yuet al.2018). Such information on crop patterns is critical for food security,and agrarian policies that are relevant to rural investment and development (Songet al.2017).In addition,the spatiotemporal dynamics of crop planting patterns have great implications for understanding the changing climate and the overall agricultural response to environmental issues (Yuet al.2014).

Four models,M3,MIRCA,GAEZ and SPAM,have been used to produce global crop distribution maps(Andersonet al.2015). When reliable input data are available,these models can generate the global spatial distribution of various crop types within croplands.However,these global models have some limitations.The maps produced by these models focus only on a specific year. The relatively coarse resolution of the maps produced by these models also limits their application at the regional level (Songet al.2017). As a result,highly accurate and long-term datasets on crop planting patterns are rare,which directly limits our ability to explore changes in crop patterns. Thus,the development of a new model in a spatially explicit manner to better understand the key processes driving crop planting pattern changes is required (Verburget al.2019). To fill this gap,this study proposed a new model to effectively spatialize statistical data to generate a time-series dataset of crop patterns at a grid. This method is used to solve the problem of agriculture land system research where there is no spatial information in agricultural statistics data and insufficient crop spatial pattern data.

2.Materials and methods

2.1.Conceptual model

The idea underpinning this model is that the crop planting pattern change in a regional area is driven by the demand for crop planting,and the distribution pattern of crop planting in this area is always in a dynamic equilibrium with the physical geography and economy of the area.The distribution of individual crops for a particular gridded cell is directly linked to the physical geography and socioeconomic factors of the cropland in the cell. Factors such as topography,temperature,precipitation,soil,population,GDP and agricultural policy vary across time and space,therefore influencing the spatial distributions of crops and driving distribution changes over time. The driving factors controlling crop distribution should be influential and representative of the research objective,and the autocorrelation among these factors should be avoided or at least controlled at a minimal level. In addition,the modeling approach proposed here hypothesizes that crop area at the regional scale,represented by statistical data,is the sum of the crop distribution at the gridded cell level. Thus,by establishing the relationship between crop planting pattern and the driving factors at the cell scale,it is possible to spatially allocate crop area to individual gridded cells and track the changes. The size of the grid is determined by the basic data that describe the crop spatial distribution. Each grid contains only one type of crop,and the sum of the area of each crop type at the grid scale must be consistent with the regional area given by the statistical data. Based on this concept,a crop spatial distribution model was developed and is shown in Fig.1.This approach was based on two interlinked modules,namely,the land-use simulation module (1st layer) and the crop pattern simulation module (2nd layer).

Fig.1 Framework visualization of the land-use and crop pattern simulation model.

The 1st layer module simulates the land-use pattern dynamics. Logistic regression is used to analyze the relationship between geographical and socioeconomic factors and the land use distribution to determine the local suitability of each land use type within a grid. The land-use demand determines the area of each land-use type in each year and is estimated from statistical data,economic simulations,or forecasting models. Based on spatial suitability and land-use demand data (statistical data),an iterative spatial allocation method determines the spatial allocation of each land-use type. The land-use simulation module realizes the spatialization of statistical data (land-use demand) and generates the cropland distribution map. The spatial distribution of cropland from the 1st layer is then inputted into the 2nd layer module to constrain the spatial allocation of individual crop types. The crop pattern simulation module uses the same method of statistical spatialization to map the crop area patterns within the cropland as determined by the 1st layer. Using the prior analysis of the driving factors that impact crop planting probabilities,the spatial iterative allocation is used to distribute the individual crop types.The two simulation modules are closely linked and are combined to complete the simulation process.

The spatial distribution model is central to both the land-use and crop pattern simulation modules. In the allocation module,a land-use type or crop type is assigned to a grid cell according to unmet statistical demand and the cell’s total probability value for that land use or crop. For each grid celli,the total probability(TPROPi,u) is calculated for each land use type or crop type using eq.(1):

wherePi,uis the suitability of locationifor the land use or crop typeu(based on the logit model),ELASuis the conversion elasticity for the land use or crop typeu,andITERuis an iteration variable specific to the land use or crop type,u.

The most important parameter for the total probability(TPROP) is the suitability of the location for a given crop type on cropland. Initially,the TPROP needs to determine the probability of the research object,Pi,that cropland for each crop appears in each grid. The accuracy of estimating the suitability of the location for a land use type or crop type directly determines the accuracy of the statistical data spatialization. Binary logistic regression is used to quantify the influence of each driver on land use and crop distributions and to determinePi. The landuse (or crop) pattern simulation module assigns land uses(or crops) to meet the area demand for each land-use (or crop) using a spatial probability function (eq.(2)):

wherePiis the probability that the grid cell at locationihas a particular land use (or crop),Xniis thenth driving factor andβnis the coefficient of the factor. The ROC characteristic is a measure for the goodness of fit of the logistic regression model similar to theR2statistic in ordinary least square regression (Pontius and Schneider 2001). The degree of fit of the regression equation for each land use (or crop) type can be tested with the ROC curve. According to the size of the area under the curve,it is judged whether the calculated land use (or crop) type probability distribution pattern has a high consistency with the real land use (or crop) type distribution.

Secondly,a practical sequence of land use or crop type conversions for the study area is estimated from the predefined conversion elasticity (ELASu) values.Defining the elasticity (ELASu) for land use or crop type conversions should fit the real situation of the research area. The relative elasticity ranges between 0 and 1.The higher the defined elasticity,the more difficult it is to cover this research type (land use or crop type). The final elasticity is calibrated by debugging the model.

Finally,using the total probabilities,the driver constraints and the area demands (statistical data),the allocation module performs multiple spatial iterations(Fig.2). The model sets the initial value ofITERuwhen the allocation module starts running and then iteratively allocates space for statistical data. The model continuously adjusts theITERuvalue according to the completion percentage of the allocation of the statistical data. The iterative spatial allocation process has three steps. Firstly,the allocation module temporarily assigns the same iterative variable (ITERu) value to each crop type and allocates the crop type with the highest total cell probability to each grid cell. Secondly,the initial allocation area of each crop type is compared with the area demand of each crop. In this step,if the area of a crop allocated in the first step is smaller than its area demand,then theITERuvalue is increased. Otherwise,theITERuvalue is decreased. Thirdly,the module conducts a second spatial crop distribution. Steps 2 and 3 are repeated until the spatially distributed area of each crop is equal to its demand. The module thus completes the spatialization of the crop demand data.

Fig.2 Visualization of the space iterative allocation process for statistical data.

2.2.Study area

We tested the proposed method in Heilongjiang Province in northeastern China (Fig.3). Heilongjiang Province has an area of 470 000 ha spanning longitudinally from 121°13′E to 135°05′E and latitudinally from 43°22′E to 53°24′N (Gao and Liu 2011);it is an agriculturally important region within China (Yaoet al.2015) and has 160 000 ha of cropland. There are two mountain areas (Daxing’anling and Xiaoxing’anling) in the northern and northwestern regions and two plains (Songnen and Sanjiang) in the western and eastern regions of Heilongjiang. The climate is temperate humid or subhumid continental monsoon.The mean annual temperature ranges from ?5 to 5°C,with an average maximum temperature of 21-22°C in July and an average minimum temperature of ?18°C in January. The mean annual precipitation is 500-650 mm,and 80% of rainfall occurs between May and September.The frost-free period is approximately 100-150 days (Chenet al.2012;Liet al.2012). In Heilongjiang,the per capita cropland area is approximately 0.31 ha,which is higher than the national level. Heilongjiang Province is divided into four agricultural zone divisions:District I,Xing’anling Mountains Forest Area (XMFA);District II,Sanjiang Plain(SJP);District III,Songnen Plain (SNP);and District IV,Laoye &Zhangguangcai Mountain Agriculture Area (LZAA).Grain production reached 6.24×1010kg in Heilongjiang Province in 2014,which is a 2.38×109kg (3.81%)increase over previous years. The total grain export from Heilongjiang is 10% of the grain yield in China. The province is one of China’s most important food production areas (Xiaet al.2014).

Fig.3 Land-use maps of Heilongjiang Province,China for the years 2000,2005,2010,2015 and 2019. I,Xing’anling Mountains Forest Area;II,Sanjiang Plain;III,Songnen Plain;IV,Laoye &Zhangguangcai Mountain Agriculture Area.

2.3.Data sources

The input data used in the crop spatial distribution model was shown in Table 1. A total of 14 driving factors,including biophysical and socioeconomic variables,were used in this study.

The data were reprocessed and standardized to ensure that data from different sources had the same spatiotemporal resolution. All the spatial data were converted to a GIS grid with a 1 km×1 km cell size and were stored in the ASCII raster format.

3.Results

3.1.Land-use pattern simulation

For each land-use type,logistic regression was used to analyze the relationships between the land-use type and its driving factors and to estimate the values of the factor coefficients. The statistically significant driving factors are summarized in Table 2. Overall,the spatial distribution of the seven land-use types was acceptably explained by the selected drivers,as indicated by the values of the receiving operator characteristic (ROC) statistic (Table 2).The value of ROC is 0.81-0.98,indicating that the regression equations constructed by all land use types and driving factors have a high degree of fit,and the probability distribution pattern of the research objects is highly consistent with the actual situation.

Table 1 Details of the input datasets

Table 2 Land-use probability coefficients (β values) in Heilongjiang Province,China derived by logistic regression1)

The built-up area,forest,and unused land had the highest ROC values of 0.98,0.96,and 0.96,respectively.Built-up areas are significantly linked to air temperature and per-capita GDP,while forest appears in areas far from economic and population centers,generally with high elevation or on sloping terrain that is moderately well drained. Unused land is found in regions with poor soil texture and poor drainage. Water bodies are negatively correlated with elevation,slope,population,and temperature. The regression results indicate that altitude,slope,and population are significantly linked to the spatial distribution of croplands (Zhanget al.2011;Chenet al.2013). Relatively low ROC values,0.87 and 0.81,were obtained for wetlands and grasslands,respectively. These results show that the logistic equation can effectively model the spatial distribution of the various land uses,and it is used in the spatial allocation module to calculate the probability (Pi) of each land-use type in each grid cell.Then,the conversion elasticity (ELASu) values of each land use type are set according to the actual situation of land use in Heilongjiang. After setting the initial value of ITER,the allocation module performs spatial iterative allocation based on the TPROP of each land use type.

This study selected 2000 as the initial year for the landuse change simulation. Land-use spatial data from 2000 and demand data (land use area statistics) were input into the allocation module (1st layer),and the cells were iteratively allocated to each land-use type based on the TPROP data. When the model allocated each land use type consistent with the demand data,the model stopped the allocation to generate space allocation data. This process was repeated for each year from 2000 to 2019.The simulated land-use distribution in Heilongjiang for 2000,2005,2010,2015 and 2019 was mapped in Fig.3.To facilitate the development of spatial research,this study introduces the division of agricultural areas to analyze spatial changes. The results show that the wetland area in the SJP has been shrinking and converted into cropland.At the same time,the continuous expansion of built-up land has occupied parts of cropland and forest. Builtup expansion can be clearly seen in SNP and LZAA,especially in several major cities. The grassland area in XMFA has been steadily decreasing,and some of the grassland has been converted into forest or built-up areas.

This study adopted a two-layer nested operation mode;the 1st layer of the land use allocation simulation was input to determine the 2nd layer in the crop allocation simulation. The 1st layer of cropland spatiotemporal data was used to control the range of the crop pattern distribution in the 2nd layer. Therefore,the annual cropland spatial data were input into the 2nd layer allocation module to determine the annual crop planting spatial pattern allocation.

3.2.Crop planting pattern simulation

Before the crop (2nd layer) planting pattern simulation,an analysis of the crop driving factors is required. The logistic regression coefficients of the drivers for each crop type in Heilongjiang are listed in Table 3. Relatively high ROC values were obtained from the regression analysis,ranging from 0.82 to 0.92. The rice distribution is closely linked to the slope,temperature,and soil drainage. The selection of driving factors affecting crop planting patterns is different from the land use patterns. More factors affecting crop planting are considered,such as some soil type data,for analysis in this study. These results make it clear that the spatialization of statistical data can be applied effectively to simulate regional spatial crop patterns. The total probability equation used in this logistic analysis accurately explains the spatial distribution of crops. Individual crops must reside within cropland;therefore,the cropland of the 1st layer simulation(land use) is input into the crop simulation module(2nd layer). Using the total probability estimates,the model can spatially allocate the total cropland and thensimulate the individual crop distributions. The crop simulation used 2005 remote sensing identification data from Heilongjiang. This simulation distinguishes three crops:rice,maize and soybean. The total probability(TPROP) of each crop type was used to perform spatial iterative allocation of the crop planting area (statistical data) based on the crop spatial data from Heilongjiang Province in 2005. Through the spatial allocation of the annual crop planting area,the three crop planting structures were simulated in Heilongjiang Province from 2005 to 2109. The simulated crop planting pattern in Heilongjiang in 2005,2010,2015 and 2019 was shown in Fig.4. Rice expansion was the highest in the northern part of SJP,especially from 2005 to 2010,after which the growth in the rice crop area in the region slowed. At the same time,there was a small increase in rice planting in the western SNP region. The planting area of soybean crops has been increasing year by year,especially in the northern part of SNP. The area of soybean in other areas changed little. Maize was predominantly concentrated within central SNP. As its area grew,maize was also planted in western SJP.

Fig.4 Simulated crop planting pattern maps of Heilongjiang Province,China for the years 2005,2010,2015 and 2019. I,Xing’anling Mountains Forest Area;II,Sanjiang Plain;III,Songnen Plain;IV,Laoye &Zhangguangcai Mountain Agriculture Area.

Table 3 Beta values of the spatial distribution of the crop patterns in Heilongjiang,China derived from logistic regression1)

3.3.Validation

To verify the accuracy of the model,the spatial allocation results for the land use and crop types needed to be validated with remote sensing image interpretation.Adjustments to the conversion elasticity (ELASu)values of the model according to the results of the accuracy validation and iteratively allocating space were continuously performed to realize spatialization of the statistical data. The land-use and crop planting pattern maps derived from visually interpreted remote sensing images were used for accuracy assessment. This study selected a year close to the simulation target and combined it with the available verification data,landuse data from 2015 and crop data from 2011,to validate the accuracy of the simulation. Since the accuracy of the cropland areas simulated in the 1st layer model directly affects the simulation accuracy of the second layer,separate accuracy analyses for cropland and crop simulations were conducted.

This study employed two methods to test the accuracy of model simulation. One method is the confusion matrix for accuracy assessment. Viewing remote sensing data as the ground truth,a mismatched grid cell between model simulation and remote sensing data was identified as a misclassified cell by the model (Liet al.2014). Another method is the simple random sampling method. In total,62 876 sample points in the research area of Heilongjiang Province were randomly selected to evaluate the accuracy of the model. A confusion matrix was constructed by identifying whether a grid cell at each of the sampling points was correctly classified or misclassified by the model. The overall accuracy,producer’s accuracy,and user’s accuracy (Table 4)in the confusion matrix are important parameters for the accuracy assessment of model simulation results(Olofssonet al.2014). The overall accuracy of cropland and crops map was 88.11 and 81.17%. The results show that the producer accuracy and the user accuracy of cropland are both evaluated at the level of 90%. The simulation accuracy of crop planting patterns is a little lower than cropland. The producer accuracy and user accuracy of rice are 76.03 and 74.27%,respectively,and those of corn are 84.24 and 84.45%,respectively.By comparing simulation results and remote sensing classification data,it is found (Fig.5) that the abnormal values of crop planting patterns mainly appear in cropping interlaced areas. The influencing mechanism of crop planting in these areas is relatively complicated,and thus is more difficult to simulate. The results show that the simulation accuracy of cropland and crop planting patterns is good,and the two-layer nested model simulation of this research can be realized.

Fig.5 Validation of the simulated maps and remote sensing interpretation maps.

Table 4 Model simulation accuracy analysis

The other method is a multiple resolution procedure for model goodness of fit. The method of an expanding detection window was used to gradually reduce the comparison resolution to test the accuracy of thesimulation data. As a measurement at one resolution is not sufficient to describe complex patterns,the objective method of Costanza (1989) was used to validate the proposed simulation model,which measures the goodness of fit of the model according to a multiple resolution procedure. The fit of each sampling window was estimated as 1 minus the proportion of cells that would have to change for the sampling windows to each have the same number of cells in each category,regardless of their spatial arrangement. This validation method showed that the goodness of fit of the model for the spatialized cropland and crop statistical areas were 0.92 and 0.91,respectively. Some of the physical farming conditions in Heilongjiang may have prevented higher accuracies. Resampling from 250 m to 1 km may have introduced additional sources of error. Nevertheless,the assessed accuracy level was sufficient to support subsequent analyses.

4.Discussion

4.1.Modeling approach

This paper outlines a novel approach for the spatialization of statistical data for the simulation of dynamic crop changes at the regional level. Using this method,statistical data without explicit spatial information can be spatialized based on geographic characteristics. Remote sensing has proven to be an important tool for crop pattern mapping (Kuenzer and Knauer 2013). However,it is difficult to interpret complex crop planting patterns using remote sensing images. It is even more challenging to generate long time-series data. A model-based crop mapping approach has advantages over remote sensingbased approaches. One advantage is that models can help to understand the processes that drive crop pattern changes. Existing models such as SPAM and GAEZ mainly focus on simulating dynamic land-use changes but pay little attention to specific crop types. In the present study,a new approach to simulate changes in crop planting patterns at the local scale using spatialization of statistical data has been applied to Heilongjiang Province.Another advantage is the ability to simulate crop patterns in different scenarios in the future. With the addition of external mathematical models to predict the planting area of crops,this method can effectively predict future crop type patterns.

The model is further improved based on the CLUE-S model’s spatial allocation module (Verburget al.2002).As a result,a two-layer nested model was developed on top of the one-layer model of CLUE-S. While the CLUE-S model can only be used to model the spatial processes of land use change,our new model has improved capacity of dynamic spatial simulation of the spatial-temporal patterns of specific crops within the cropland. The proposed method simulates distribution changes in two stages:the 1st layer simulates the spatial distribution of cropland and the 2nd layer simulates crop type distributions within the cropland. This approach can improve simulation accuracies and reduce the error in crop spatial simulation. Logistic regression analysis of the spatial dynamics of the land use types and the crop types supported a spatial-statistics land-use model and spatial allocation of the crop area. All the possible driving factors ranging from natural geographic to socioeconomic factors were considered. Simulating crop planting patterns after the prior generation of a cropland layer improved the simulation accuracy. Moreover,a spatial iteration method based on multivariate spatial correlation was incorporated into the modeling approach to spatially allocate the known crop area demand into the underlying grids. Two innovations make the proposed method superior to other methods for producing long time-series crop maps. First,the accuracy of the spatial allocation was improved by analyzing the influence of spatial correlations with multiple driving factors. Second,the total demand of the crop type area was controlled by spatial iterations to ensure that the total allocated area agreed with known statistics.

However,the proposed method may have certain limitations. The spatial simulation of crop type is based on the determination of the TPROP and the suitability of location (P) values. While the logistic analysis can identify the drivers of changes in crop patterns,if one or several drivers have extreme changes or if the crop type pattern is strongly affected by human factors,then the model cannot simulate the spatial changes. Future research should further determine the influences of cropping decisions and allow timely linkage to the processes affecting the spatial pattern of crops with their driving factors.

4.2.Implications

The simulations effectively reproduced changes in crop planting patterns between 2005 and 2019 in Heilongjiang,demonstrating the capacity of the spatialization of statistical data to generate multitemporal crop maps.However,some issues remain for future research.

First,remote sensing-derived crop pattern maps were used as reference data to simulate more recent changes in crop patterns. The accuracy of the reference data therefore has a considerable impact on the accuracy of the modeling. The reference maps we used were resampled from a 250-m spatial resolution to a 1-km resolution to match other data sources. Although resampling should ensure data consistency,additional errors are inevitable,and these errors persist throughout the simulation. Simulation accuracy could be improved by using only data with a consistent spatial resolution to avoid resampling. A hard classification was used to allocate the crop area demand for specific grids. The model spatially distributed the area demand to specific grids using an iterative method,which ensured that the total area of each crop type was correct but also limited the spatial distribution range.

Second,the method for the spatialization of statistical data simulated changes in crop spatial patterns from 2005 to 2019 based on validation using crop patterns in 2015.In this study,spatial data are needed for both spatial suitability analysis of the research object and validation of the accuracy of the allocation results. However,due to the lack of spatial data on crop planting in 2000,the starting year of the spatial allocation of the crop planting area in the 2nd layer does not completely coincide with the 1st layer. Although this method only simulates the crop planting pattern from 2005 to 2019,the research results can confirm that the statistical data spatial allocation method of this research is effective. If data on the spatial pattern of crop planting in 2000 could be obtained,then this study could conduct related research.Additionally,the model is equally capable of simulating future crop pattern scenarios. In the future,we will try to simulate the spatiotemporal pattern of crops by setting different target scenarios to further verify the ability of this method to simulate future scenarios.

Third,the selection of driving factors needs to be further improved in future research. This study selects some of the commonly used driving factors to analyze the impact of crop spatial pattern. As a precaution to control autocorrelation,only the most influential factor is selected in this analysis. Agricultural policy factors are predominantly reflected in specific areas,such as basic farmland protected areas and ecological protection restricted areas. The research objects do not change into other land use types in restricted areas. At the same time,the area dynamics of crops caused by agricultural policies can be used to reflect the changes in the area of crops spatially (Zhouet al.2020). For example,the state has adopted policies to increase rice planting in Heilongjiang,and the proposed method can determine the spatial planting area of these newly added rice crops in the future. However,this study still has some shortcomings in considering the influence of policy factors. Because it is not easy to use quantitative analysis for policy factors,this study does not directly apply policy factors to the driving factors for the object research. However,to reflect the influence of policy factors on research,this study indirectly reflects the influence of policies on some socioeconomic factors. Future research can try to quantify policies and directly input them as driving factors for analysis.

5.Conclusion

A method for the spatialization of statistical data,a novel spatial allocation module with a multilevel structure,was proposed in this paper. In this study,a two-layer nested structure of the statistical data spatialization method was adopted to realize the spatial distribution of the crop planting area within cropland based on the spatial distribution of land use area data. In addition,this method was used to simulate land use and crop planting patterns in Heilongjiang Province from 2000 to 2019. The spatialization method simulated the distribution of land use types and crop types to grid cells by integrating natural geographic and socioeconomic data with land-use and crop maps derived from remote sensing images. The remote sensing interpretation results were used to verify the accuracy of the statistical data spatialization results,and the results showed that the accuracy of the distribution of cropland and crop patterns was 91.8 and 91.4%,respectively. Our results suggest that the proposed model is a suitable tool for exploring the spatial characteristics of natural resources and agricultural production.

The exploration and application of the spatialization of crop statistical data are powerful supplements to the remote sensing interpretation method,which solves the need for long-term and large-scale crop spatial patterns.At the same time,this method can also fulfill the demand for crop spatial pattern simulation under different demand scenarios in the future. The described model is effective not only in forecasting future crop patterns but also in simulating long time-series crop planting patterns at the regional scale.

Acknowledgements

The research described in this paper is supported and financed by the National key Research and Development Program of China (2019YFA0607400),and the Fundamental Research Funds for the Central Universities,China (CCNU19TS045). All persons and institutes who kindly made their data available for this analysis are acknowledged.

Declaration of competing interest

The authors declare that they have no conflict of interest.