Quantifying the agreement and accuracy characteristics of four satellite-based LULC products for cropland classification in China

Jie Xue ,Xianglin Zhang ,Songchao Chen, ,Bifeng Hu ,Nan Wang,Zhou Shi,

1 Department of Land Management,Zhejiang University,Hangzhou 310058,China

2 Institute of Applied Remote Sensing and Information Technology,College of Environmental and Resource Sciences,Zhejiang University,Hangzhou 310058,China

3 ZJU-Hangzhou Global Scientific and Technological Innovation Center,Zhejiang University,Hangzhou 311215,China

4 Department of Land Resource Management,School of Public Finance and Public Administration,Jiangxi University of Finance and Economics,Nanchang 330013,China

5 Key Laboratory of Data Science in Finance and Economics,Jiangxi University of Finance and Economics,Nanchang 330013,China

6 Key Laboratory of Spectroscopy Sensing,Ministry of Agriculture and Rural Affairs,Hangzhou 310058,China

Abstract Various land use and land cover (LULC) products have been produced over the past decade with the development of remote sensing technology.Despite the differences in LULC classification schemes,there is a lack of research on assessing the accuracy of their application to croplands in a unified framework.Thus,this study evaluated the spatial and area accuracies of cropland classification for four commonly used global LULC products (i.e.,MCD12Q1 V6,GlobCover2009,FROM-GLC and GlobeLand30) based on the harmonised FAO criterion,and quantified the relationships between four factors (i.e.,slope,elevation,field size and crop system) and cropland classification agreement.The validation results indicated that MCD12Q1 and GlobeLand30 performed well in cropland classification regarding spatial consistency,with overall accuracies of 94.90 and 93.52%,respectively.The FROMGLC showed the worst performance,with an overall accuracy of 83.17%.Overlaying the cropland generated by the four global LULC products,we found the proportions of complete agreement and disagreement were 15.51 and 44.72% for the cropland classification,respectively.High consistency was mainly observed in the Northeast China Plain,the Huang-Huai-Hai Plain and the northern part of the Middle-lower Yangtze Plain,China.In contrast,low consistency was detected primarily on the eastern edge of the northern and semiarid region,the Yunnan-Guizhou Plateau and southern China.Field size was the most important factor for mapping cropland.For area accuracy,compared with China Statistical Yearbook data at the provincial scale,the accuracies of different products in descending order were: GlobeLand30,FROM-GLC,MCD12Q1,and GlobCover2009.The cropland classification schemes mainly caused large area deviations among the four products,and they also resulted in the different ranks of spatial accuracy and area accuracy among the four products.Our results can provide valuable suggestions for selecting cropland products at the national or provincial scale and help cropland mapping and reconstruction,which is essential for food security and crop management,so they can also contribute to achieving the Sustainable Development Goals issued by the United Nations.

Keywords: global LULC products,cropland mapping,accuracy evaluation,food security,China

1.Introduction

The need to satisfy the demands for food,biofuel and other commodities to support the growing human population exerts considerable pressure on global agriculture (Foleyetal.2005;Godfrayetal.2010;Tilmanetal.2011;Yu Q Yetal.2023).China is one of the largest agricultural countries,yet it confronts immense pressures to provide food for 18% of the global population with less than 9% of arable land (Zuoetal.2018;Yu Z Netal.2023).Therefore,accomplishing the 2030 Zero Hunger of Sustainable Development Goals (SDGs) for China will be a great challenge (Alexandratos and Bruinsma 2012;Fuetal.2022).As one of the most essential land use and land cover (LULC) types,cropland is the base for the survival of humans and the sustainable development of society.Therefore,accurate information on the quantity and spatial distribution of cropland in China is an absolute necessity for ensuring crop production and food security.

Remote sensing is a cost-effective and efficient tool for mapping and monitoring cropland (Grekousisetal.2015;Huetal.2021).Since the early 1990s,many researchers have endeavoured to map global or regional LULC changes using remote sensing technology.The International Geosphere Biosphere Programme Data and Information Systems Cover (IGBP-DISCover),sourced from the United States Geological Survey,was the earliest effort for land mapping,which was conducted at 1 km spatial resolution in 1992 (Loveland and Belward 1997).The Moderate Resolution Imaging Spectroradiometer (MODIS) Land Cover dataset created by Boston University,USA,as well as the Global Land Cover Map 2009 (GlobCover2009) and Climate Change Initiative Land Cover (CCI-LC) datasets from the European Space Agency,were launched in the early 21st century (Friedletal.2002;Bontempsetal.2011;Defournyetal.2016).However,these products have relatively low spatial resolutions of 300-500 m.Since Landsat satellite data became publicly available and was accompanied with the increasing power of computers,many more fine-resolution LULC products were generated,such as FROM-GLC and GlobeLand30 (Gongetal.2013;Chenetal.2015).Meanwhile,Potapovetal.(2022)derived crop maps with a spatial resolution of 30 m.Many other studies,including Yuetal.(2014),Fritzetal.(2015),and Xiongetal.(2017),have also produced cropland maps by satellite image classification of existing datasets.

However,the differences in LULC classification systems,spatial resolutions,satellite sensors and mapping approaches lead to clear differences in the cropland products (Tubielloetal.2023).Therefore,evaluating the global cropland products for China to support decisionmaking is of great significance.Many researchers have assessed the accuracy of LULC products at global or regional scales (Heroldetal.2008;Yangetal.2017;Weietal.2020).For example,Wuetal.(2008) and Ranetal.(2010) evaluated the same land cover datasets (IGBP DISCover,UMd,GLC2000 and MODIS),but the former indicated that MODIS was the best cropland dataset for China,while the latter concluded that GLC2000 showed the highest accuracy of overall land cover types among the products tested.Although the same validation database was used in these two studies,the results were inconsistent.Current research has indicated that cropland products often show great disagreements,while the underlying drivers of the inconsistencies remain poorly explored.Additionally,previous studies mainly concentrated on a time-frame of 2000,so the LULC product evaluation was at low spatial resolution and is now outdated.Meanwhile,Venteretal.(2022) recently addressed the uncertainties and challenges in product evaluation.Thus,a more detailed and up-to-date evaluation of the cropland products,and the exploration of the drivers for cropland mapping accuracy and consistency,are urgently needed.

To tackle this knowledge gap,this study mainly aims to: (1) harmonise the cropland classification schemes of four major global LULC products (MCD12Q1 V6,GlobCover2009,FROM-GLC,and GlobeLand30) based on the FAO criterion;(2) evaluate the areal and spatial consistency of these four cropland products;and (3)analyse the potential reasons for inconsistencies among the four cropland products.The results obtained in this study can help related researchers to select suitable cropland products for China,and provide new implications for improving cropland mapping.

2.Materials and methods

2.1.Global cropland products and auxiliary data

Four popular global cropland maps (MCD12Q1 V6,GlobCover2009,FROM-GLC and GlobeLand30) were used in this study,and the relevant information for them is shown in Table 1.MCD12Q1 V6 is an annual land cover product at 500 m spatial resolution from 2001 to 2020 that was produced by the decision tree classification algorithm based on MODIS reflectance data.It has six different classification schemes,and we used the “LC_Prop2” layer based on LCCS land use from the Food and Agriculture Organization (FAO)(Friedl and Sulla-Menashe 2019).GlobCover2009 is a global land cover product produced by the European Space Agency (ESA) using 2009 medium resolution imaging spectrometer (MERIS) images with a 300 m spatial resolution (Bontempsetal.2011).The FROMGLC is the first 30 m resolution global land cover map,which was generated by Tsinghua University based on Landsat images (Gongetal.2013).The National Geomatics Center of China produced GlobeLand30 2000/2010/2020,which was based on the classification approach of pixel-object-knowledge (POK) and used the Chinese environmental and disaster satellite (HJ-1)imagery and Landsat TM/ETM+images (Chenetal.2014,2015).These four products overlapped in time around 2010,so we evaluated the consistency and accuracy based on the data for 2010 in this study.

To assess the cropland area accuracy,the relevant statistical data were acquired from the FAOSTAT land use database (http://www.fao.org) andChinaStatistical Yearbook(https://www.cnki.net/).The FAO statistical data were used to evaluate the national area accuracy,while theChinaStatisticalYearbookdata were used for provincial assessments,due to the lack of sufficiently detailed data in FAO.TheChinaStatisticalYearbookdata are based on the second China land survey and China land change survey,which are the most authoritative sources for cropland data.Meanwhile,all classification schemes of the cropland products used in this study were converted to match the definition of cropland from the FAO and Chinese criteria (Table 2).

Four potential control factors (i.e.,slope,elevation,field size and crop system) were selected,and theirrelationships with the consistency of the cropland products were analysed based on previous studies (Weietal.2020;Tesfayeetal.2021;Sunetal.2022;Zhang Cetal.2022).Elevation information was acquired from the shuttle radar topographic mission (SRTM,https://www2.jpl.nasa.gov/srtm/).Slope was calculated based on elevation in SAGA GIS (Conradetal.2015).The field size dataset was collected from Google Maps and Bing using the Geo-Wiki application,which was independent of any cropland product used in this study.This application was provided by Lesivetal.(2019),and it can be downloaded at http://pure.iiasa.ac.at/id/eprint/15526/.This crop system represented the cropping times of a year for a field,which was generated by the normalised difference vegetation index from SPOT-VGT data,and was downloaded from the Resource and Environment Data Cloud Platform (http://www.resdc.cn/).

Table 1 Descriptions of the four global land use and land cover (LULC) products

2.2.Methodology

Areal evaluation of cropland productsUsing a single classification scheme is the basis for determining the comparability and compatibility of different cropland products (Heroldetal.2008).In this study,we estimated the cropland areas at the national and provincial levels for each class of the four global LULC products,and then made comparisons with FAO statistics andChina StatisticalYearbookdata separately.As shown in Table 2,we harmonised the classification schemes of the four products using the FAO criterion,which defines “cropland”as “arable land” and “land under permanent crops”.

The coefficient of determination (R2) was used to evaluate the dispersion and fitness of cropland areas between the cropland products andChinaStatistical Yearbookdata at the provincial scale.A largerR2value indicated that the cropland was better suited to the statistics.R2was calculated as:

wheresiis the cropland area estimated from the four products in provincei;tiis the cropland area fromChina StatisticalYearbookof provincei;nis the number of Chinese provinces;andis the mean cropland area of the statistical data of all Chinese provinces.

Spatial consistency and accuracy evaluation of cropland productsThe spatial evaluation included spatial consistency and accuracy,which were analysed by spatial overlay and validation samples.Spatial consistency describes whether the cropland classifications of the four global LULC products are in agreement in the same situation by overlapping these products onto a new composite map (Yangetal.2017).The pixel values ranged from 1 to 4,representing spatial consistency from the lowest to the highest degrees.A value of 1 means disagreement,where only one product identified the pixel as cropland.As the value increases,the spatial consistency between these cropland datasets increases.A value of 4 means complete agreement,indicating that all four products defined the pixel as cropland.

To analyse the spatial accuracy,we divided China into 1,101 grids of 1°×1° on Google Earth Engine,and ten validation samples were randomly generated in each grid.This sampling method ensures the objectivity of the distribution of test samples.The LULC type of each validation sample was judgedviavisual interpretation of the high-resolution imagery on Google Earth by three experienced interpreters.The imagery we chose was for the year of 2010,which is at spatial resolutions ranging from 156.3 to 0.3 m.The timeline of Google Earth can help us identify LULC classes with seasonal characteristics by reviewing images at different stages.When there was no valid imagery due to the rain or cloud,we used the images from the year before and after as an alternative for confirming the land use type.The interpreters could not change the sample locations,which means that all the samples should be labelled as either cropland or non-cropland.After interpreting the respective areas,the sample labels were further cross-validated and determined by the three interpreters.Ultimately,we interpreted 2,569 cropland samples and 7,470 noncropland samples in China (Fig.1).

Fig.1 Location of the validation samples for cropland and non-cropland throughout China in 2010.

A confusion matrix was used to calculate the overall accuracy (OA),commission error (CE),and omission error (OE) of the four datasets.These indices are defined as:

where true positive (TP) and true negative (TN) indicate the number of points classified correctly as either a positive class or a negative class;false positive (FP)means the number of negative points incorrectly classified as positive;and false negative (FN) means the number of positive points misclassified as negative.In addition,the Kappa coefficient (Kappa) was adopted for the accuracy evaluation of the LULC classification of the four products and the validation samples.It was calculated as:

where OA represents the sum of classified samples divided by the total number of samples,and peis the product of the sum of the LULC classifications of the four products and validation samples relating to each class divided by the square of the total number of pixels.

The effects of potential control factors on cropland product consistencyWe adopted the Random Forest(RF) algorithm to calculate the relative contribution of each potential control factor (i.e.,slope,elevation,field size and crop system) for mapping the cropland extent.RF based on a bagging ensemble learning algorithm creates many single predictors and randomly selected samples and features to fit each predictor (Breiman 2001).The importance of each factor was then calculated by the increased mean square error of each feature by including or eliminating tree models in the out-of-bag samples (Huetal.2020,2023;Zhang X Letal.2022).

3.Results

3.1.Classification scheme transformations

The land use classification recommended by FAO was used as the benchmark scheme for cropland classification in this study (Table 2).In the FAO scheme,there are two land use types that belong to the cropland,“arable land” and “land under permanent crops”.The arable land included “land under temporary crops”,“land under temporary meadows” and “land temporarily fallow”.Generally,land under temporary crops means the land used for crops with a growing cycle of less than one year,and it is often referred to the physical areas of land on which temporary crops are grown.Under this definition,we included “herbaceous croplands” in MCD12Q1,“post-flooding or irrigated croplands” and“rainfed croplands” in GlobCover2009,and “cropland”in FROM-GLC in this category.We matched “natural herbaceous/croplands mosaics” of MCD12Q1 to the class of land under temporary meadows,which included land temporarily cultivated with herbaceous forage crops for mowing or pasture.“Bare herbaceous croplands” of FROM-GLC are mainly defined as recently harvested and fallow land,so they corresponded to the class of land temporarily fallow.Land under permanent crops describes land that is cultivated with long-term crops which do not have to be replanted for several years,such as land under trees and shrubs producing flowers and nurseries.Therefore,“forest/cropland mosaics” in MCD12Q1,“mosaic cropland” and “mosaic vegetation/cropland” in GlobCover2009 and “orchards” in FROMGLC were included in this category.Furthermore,because there is only one class in GlobeLand30 that defines the cropland,we matched “cultivated land” in this product to cropland.

To compare the area consistencies of the cropland products at the provincial scale,we also harmonised the cropland classifications of the four products to the China national land survey scheme.They are the 36 and 35 of MCD12Q1,the 11 and 14 of GlovCover,the 10 of FROM-GLC,and the 10 of GlobeLand30.The data for the cropland area of each province were from theChina StatisticalYearbook.

3.2.Areal evaluation of the four LULC products for cropland classification

Cropland area consistency with FAO statisticsThe area comparison between the four cropland products and the FAO statistics (Fig.2) shows that the area of MCD12Q1 was very close to FAO,including the subdivision of cropland.However,there were considerable overestimations by the other three products under any category in this study.Among these,FROMGLC showed the highest overestimations,followed by GlobCover2009 and GlobeLand30.The distinctions in cropland areas indicated a close relationship with the classification schemes used in the cropland products.The definition of cropland in MCD12Q1 was based on the land use scheme of the FAO-Land cover classification system.In FROM-GLC,“cropland” and “bare herbaceous croplands” corresponded to the cropland of FAO.However,the barren cropland also included all types of land not covered by vegetation,such as lake bottoms in the dry season.This criterion was one of the key reasons why its areas were larger than the areas of FAO.As for GlobCover2009,it had four categories that matched the FAO cropland,and the two mosaic classes may lead to overestimations of the cropland areas.There was only one category that matched “cropland” in GlobeLand30,which was “cultivated land”.In addition,it was not subdivided into “arable” and “land under permanent crops”.Therefore,“cultivated land” included not only“cropland” but also the pastures for grazing.Therefore,the cropland areas in GlobeLand30 were slightly larger than those in FAO.

Fig.2 Area comparisons of the four cropland products based on the FAO scheme in China at the national scale.A,the first level of cropland classification.B,the second level of cropland classification.

Cropland area consistency with China Statistical Yearbook dataThe comparisons of each province between the individual cropland products and the statistical data in China are shown in Fig.3.The size of each point represents the difference value between the product and the statistics.The cropland areas in most of provinces were overestimated relative to the statistics(Fig.3),and most of the difference values were between 2,500×103and 7,500×103ha.The results showed that the cropland area of Qinghai in GlobCover2009 was the most overestimated among the four cropland products,while underestimations occurred most frequently in MCD12Q1.However,the cropland area was most underestimated in Heilongjiang Province,which also appeared in GlobCover2009.

Fig.3 Correlations between the areas of the four cropland datasets and China Statistical Yearbook data at the provincial scale.The size of the points means the difference value between the dataset and China Statistical Yearbook data,and the larger the difference value,the larger the points.The histogram shows the total cropland area of the statistical data (grey) and the four datasets.

TheR2values of GlobeLand30,FROM-GLC,MCD12Q1 and GlobCover2009 evaluated by the validation dataset followed the order of 0.95＞0.84＞0.73＞0.27,respectively.The best performance was given by GlobeLand30,followed by FROM-GLC and MCD12Q1.GlobCover2009 showed the poorest accuracy,corresponding to the largest difference values obtained with GlobCover2009.Overall,GlobeLand30 was the best match with theChina StatisticalYearbookdata at the provincial scale.However,GlobeLand30 also tended to overestimate values at the national scale (Fig.3).The value from FROM-GLC was overestimated by about 10,485×103ha,but its area was the closest to the statistical data.The value from MCD12Q1 was underestimated by about 12,639×103ha,which was also near the statistical data.

3.3.Spatial evaluation of the four LULC products for cropland classification

Spatial consistencyThe spatial patterns of the four cropland products are shown in Fig.4.Overall,the general distributions of cropland at the national scale were broadly similar.However,there were evident discrepancies in the spatial patterns of local regions.

Fig.4 Spatial patterns of the four global cropland products.The cropland classification is based on the FAO scheme.

To better illustrate the spatial consistency,the agreement map is shown in Fig.5.Overall,among the areas defined as cropland by at least one of the four products,the proportions of complete agreement,medium agreement,low agreement and disagreement in China were 15.51,19.12,20.65,and 44.72%,respectively.These percentages indicated a large inconsistency among the four products,and the disagreement category was the largest.Regarding the spatial pattern,good consistency was found in the Northeast China Plain,the Huang-Huai-Hai Plain and the northern part of the Middlelower Yangtze Plain.This is because the croplands in those places were integrated and continuous,and the landscape had low spatial heterogeneity.However,the croplands mainly showed low and medium consistency from the major grain-producing areas to the farmingpastoral regions located in the eastern edge of the northern and semi-arid regions.Meanwhile,patterns of fragmented and small cropland areas affected the consistency of the products in some regions,such as the Yunnan-Guizhou Plateau and southern China,and the cropland agreement was the worst in pastoral areas like the Qinghai-Tibet Plateau.In conclusion,identifying cropland was challenging when the mixture of cropland,bare land and other vegetation types dominated an area.A complicated landscape and low spatial resolution increased the difficulty of cropland identification and the disagreements among the different cropland products.

Fig.5 Spatial agreement levels of the four cropland products with the nine agricultural zones of China.NCP,Northeast China Plain;HHHP,Huang-Huai-Hai Plain;MLYP,Middle-lower Yangtze Plain;SC,southern China;LP,Loess Plateau;SCB,Sichuan Basin;YGP,Yunnan-Guizhou Plateau;NASR,northern arid and semiarid region;and QTP,Qinghai-Tibet Plateau.

Spatial accuracyBased on the large number of validation samples,a confusion matrix was established to analyse the spatial accuracy of the four cropland products(Table 3).The OA values of all products were more than 80%,with MCD12Q1 having the highest OA of 94.90%,followed by GlobeLand30 (93.52%),GlobCover2009(87.96%) and FROM-GLC (83.17%).Meanwhile,the Kappa coefficients follow the same ranking: MCD12Q1(0.786)＞GlobeLand30 (0.758)＞GlobCover2009(0.581)＞FROM-GLC (0.472).

The spatial distribution of the inconsistencies between validation samples and the corresponding pixels of each cropland product is shown in Fig.6.Clearly FROM-GLC had the largest amount of error pixels,which numbered 1 768 in total.The lowest number of errors was detected in MCD12Q1,with only 606 misclassifications.As for omission errors,we found that the herbaceous type was the easiest to confuse with cropland.There were 216,258,129 and 140 cropland samples misclassified as herbaceous in MCD12Q1,GlobCover2009,FROM-GLC and GlobeLand30,respectively.Forest was the second most misclassified type,in which 132,150,188 and 140 cropland samples were incorrectly classified as forest in the abovementioned products.Small numbers of pixels were classified as water bodies,shrubs and urban and built-up lands.Meanwhile,the incorrectly classified pixels frequently occurred in the places where the four products were in high disagreement (Fig.5).

Fig.6 The disagreements between validation samples and their corresponding pixels in each cropland product.Red points represent cases where real non-cropland was classified as cropland,corresponding to commission errors;and the other coloured points indicate that the actual cropland pixels were labelled as non-cropland (e.g.,herbaceous (bright green),forest (dark green),barren (brown),water bodies (blue),shrub (celadon) and urban and built-up lands (pink)),corresponding to omission errors.

4.Discussion

4.1.Drivers of the inconsistencies and accuracy discrepancies of four global LULC products for cropland classification

Several reasons may contribute to the discrepancies in the cropland locations and extents of the different global LULC products.Firstly,the classification system and the definition of cropland strongly affected the product accuracy (Yangetal.2017;Zhangetal.2019;Gaoetal.2020).To compare the four cropland products at the national scale,we harmonised the different classification schemes to the FAO system.Therefore,the detailed class maps were transformed into a generalised scheme with fewer classes.For example,we considered every class that was related to cropland,including the mosaic class.GlobCover2009 used the UN-LCCS classification system,which has two mixed categories for the land under permanent crops.Thus,the cropland area of GlobCover2009 was vastly overestimated.Meanwhile,the classification system for the LC_Prop2 layer of MCD12Q1 was aligned with FAO,so the area of MCD12Q1 matched perfectly with the FAO statistics at the national scale.In order to compare the four cropland products based onChinaStatisticalYearbookdata at the provincial scale,we unified the products according to the Chinese definition of cropland.GlobeLand30 and FROM-GLC were produced by Chinese institutions,so their cropland areas were more closely matched to theChinaStatisticalYearbookdata than those of the other two products.

Secondly,the various classification methods and strategies can contribute to the discrepancies (Verburgetal.2011;Masetal.2014).Many researchers believe that the spatial resolution of imagery greatly impacts cropland mapping accuracy (Hansenetal.2000;Banetal.2015).However,in this study,the accuracy of MCD12Q1 at 500 m spatial resolution was higher than the other three products at 300 and 30 m resolution,indicating that the appropriate classification methods still can achieve an ideal result,despite having lower spatial resolution.As for MCD12Q1,the land use was classified using the decision tree method (Friedletal.2010).At 30 m spatial resolution,the spatial accuracies of FROM-GLC and GlobeLand30 differed widely.The former emphasized automatic computer classification and adopted the single Random Forest method,which cannot reflect the diversity land cover types in regions with complex spectral information (Yuetal.2013).The area of FROM-GLC was overestimated (Fig.2) and there were many discrepancy points of FROM-GLC in Fig.6.Thus,its accuracy was the lowest among the four cropland products.In contrast,GlobeLand30 combined automatic classification with the POK method,which added manual verification and knowledge-based modifications.Therefore,its spatial accuracy was greatly improved.However,GlobeLand30 divides the earth into equal map sheets,so the deviation of its areas was larger than the FAO statistics.GlobCover2009 combined supervised and unsupervised methods by dividing the earth into 22 separate climatic regions,so it considered the spatial heterogeneity of the same land cover type which contributed to the finer classification.Therefore,the spatial accuracy of GlobCover2009 was better than FROM-GLC.

Thirdly,classification accuracy may be strongly affected by the selection and spatial distribution of the validation data and the accuracy of training data labels(Stehman and Foody 2019;Pengraetal.2020;Stehmanetal.2021).The validation samples for GlobCover2009,FROM-GLC and GlobeLand30 were collected by visual interpretationviahigh-resolution images or field surveys across the globe,while the decision tree method adopted in MCD12Q1 did not use training samples.All of these differences would lead to variations in accuracy among the different cropland products.

Finally,some other factors could also affect the accuracy performance inconsistently.For example,the dates of the original data used to produce each of the four products may cause disparities (Zhouetal.2019;Zhang Cetal.2022).Due to fog,rain,clouds and other factors,there are often only a few images available for valid observations.Therefore,the images from previous or subsequent years were adopted as substitutes.For example,FROM-GLC used available imagery before 2006 as the alternative for places with no suitable images after 2006.However,the cropland may have been transformed to other land uses during that time,resulting in the misclassifications in these products.

4.2.The effects of potential control factors on the consistency of the various products

The main contribution of the current study is introducing a method to incorporate the effects of multiple variables on cropland mapping.A previous study has proven that the consistency of cropland is highly related to factors such as topography and farming regime (Luetal.2016).Thus,we selected slope,elevation,field size and crop system in this study for analysing the relationships between these factors and cropland consistency.This more comprehensive approach differs from previous studies,which only analysed one of the factors.

The field size is the most important factor for mapping cropland (Fig.7).The field size can reflect a heterogeneous and fragmented landscape that is distributed in various regions.Meanwhile,the correlations between each factor and cropland agreement verified the importance of field size for mapping cropland (Table 4).Evidently,the cropland agreement level declined with increases in elevation and slope,which was in line with the findings of Zhang Cetal.(2022).In addition,the consistency increased with an increase in field size,which was also indicated by the fact that the areas of complete cropland agreement were mainly distributed in regions with extensive cropland and flat terrain.The areas of cropland disagreement are distributed in regions with varying topography,where the fields are often fragmented and small.These results were comparable to the findings of previous research(Weietal.2020).Meanwhile,the inconsistent areas often confuse cropland with the herbaceous and forest types because the fragmented and small fields caused many mixed pixels (Sunetal.2022).

Fig.7 The relative importance of four factors for mapping the cropland extent.

4.3.Perspectives and limitations

Table 4 Spearman correlations between cropland consistency and the control factors

Our results indicate that MCD12Q1 is a good choice for crop map data selection at the national scale when there is not a requirement for high spatial resolution.GlobeLand30 performed well at the provincial scale and it also has relatively fine spatial resolution,though its total area was the most seriously overestimated.The main challenges for cropland mappingviaremote sensing in China are land fragmentation,and complex and diverse topography.Smallholder farms dominate the Chinese agricultural sector,especially in southern China,which causes the fields to become small and fragmented(Duanetal.2021).In northern and northeastern China,agriculture has realised large-scale farming,so the fragmented fields mainly occurred in southern China.Meanwhile,the cloudy and rainy climatic conditions affected cropland mapping as well;thus,the cropland accuracy was relatively low in southern China.

Achieving higher spatial resolution is the key to improving cropland mapping accuracy.The spatial and area comparisons showed that GlobeLand30 was able to achieve high accuracy,which was at 30 m spatial resolution.Fusing the open archive satellite images by machine learning is an effective way to achieve cropland mapping with high accuracy,such as harmonising Landsat 5,7 and 8 at 30 m spatial resolution with Sentinel-2 at 10 m spatial resolution (Gongetal.2019;Zhangetal.2020;Yang and Huang 2021;Chenetal.2022).Meanwhile,the validation samples have been shared in some open platforms in recent years (e.g.,Geo-wiki),which makes cropland mapping more convenient (Fritzetal.2012).However,more time series and higherquality samples need to be shared for high accuracy cropland or other LULC mapping applications.

This study mainly focused on evaluating the performance of the main global LULC datasets in China.However,there are some other China-specific LULC products,such as China’s Land Use/Cover Dataset(CLUD) (Liuetal.2014) and China Land Cover Dataset(CLCD) (Yang and Huang 2021).In future studies,we will unify the evaluation framework of global and national datasets to explore the effects of the various classifier strategies at different scales.

5.Conclusion

The comprehensive assessment and comparison analysis of different cropland products can provide a significant reference for data feasibility and selection.In this study,four cropland products (MCD12Q1 V6,GlobCover2009,FROM-GLC and GlobeLand30) were assessed for areal and spatial consistency and accuracy based on the FAO classification scheme and using the data for China circa 2010.The influencing factors were analysed as well.At the national scale,MCD12Q1 showed the best area consistency using the FAO scheme,with overall spatial accuracy of 94.90%,followed by GlobeLand30,GlobCover2009 and FROM-GLC.Compared with theChinaStatisticalYearbookdata at the provincial scale,GlobeLand30 performed best (R2=0.95),followed by FROM-GLC,MCD12Q1 V6 and GlobCover2009.GlobCover2009 showed a great amount of deviation from the statistics.Regarding the spatial agreement,the disagreement among the four cropland products reached 44.72%.Field size was the most important driver influencing the accuracy of cropland mapping,and the results showed that a smaller field size led to lower cropland consistency.Explicit and precise information on the cropland extent and location is essential for achieving sustainable development goals,especially in China,and is also fundamental for food security assessment and management.Evaluating the performance and accuracy of cropland products is the key to using them appropriately.

Acknowledgements

This study was supported by the National Key Research and Development Program of China (2022YFB3903503),the National Natural Science Foundation of China(U1901601),and the Science and Technology Project of the Department of Education of Jiangxi Province,China(GJJ210541).

Declaration of competing interests

The authors declare that they have no conflict of interest.