Abstract
Cropland cover datasets is of great significance for research on agricultural monitoring. The existing investigations on the inconsistency of different cropland datasets have mainly focused on first-class cropland and only analyzed the causes of this inconsistency in terms of cartography. To date, investigations have neglected the importance of fine cropland types in studies such as global food security assessment, and a comprehensive analysis of the causes of inconsistency from the perspectives of both cartography and geography is lacking. Moreover, the verification samples of existing studies have primarily been collected based on Google Earth. So, we examined the cropland resources of Cambodia using areal, spatial consistency, elevation classification, and field survey data assessment methods for the Global Food Security-support Analysis Data at 30 m for Southeast Asia, Global Land Cover Fine Surface Cover30-2015, Finer Resolution Observation and Monitoring of Global Land Cover2015, and SERVIR-Mekong datasets and comprehensively investigated the causes of inconsistency in terms of geography and cartography. The results revealed that the consistency of the extracted areas of first-class cropland among the four datasets was high. But, the cropland areas and statistical results from the Food and Agriculture Organization (FAO) of the United Nations are quite different. The overall accuracy (OA) for the first-class cropland of GFSAD30SEACE, GLC_FCS30-2015, and SERVIR-Mekong datasets were >82%. For fine cropland types, however, the OA of the SERVIR-Mekong dataset was relatively high, at 74.87%, while the accuracy levels of the global-scale GLC_FCS30-2015 and FROM_GLC2015 datasets were <50% due to the influence of scale size on mapping accuracy. In addition, in the eastern and northern portions of Cambodia with elevations of 50–200 m, the spatial consistency of the four datasets was low due to the serious confusion between cropland and forest, grassland, and shrub types. Therefore, land cover producers should adopt a zonal stratification strategy, focusing on remote sensing extraction techniques for confusing types in areas with high inconsistency to improve the accuracy of cropland.
1 Introduction
Cropland, as an important type of land surface cover, is the basis for the survival and development of human society. The area, quantity, state, and function (such as production function) of cropland will change continuously with time, and these variations play an important role in human social and economic development, food security, and global climate change [1,2,3,4]. Globally, cropland is becoming a scarce resource, indicating the need for more efficient land-use distribution and agricultural innovation in order to improve global land-use efficiency [5]. To address issues such as protecting forest ecosystems while increasing food production in sustainable development, achieving a transition to more environmentally friendly and sustainable land use, and formulating policies that coordinate development with nature protection [6,7], it is necessary to master the use of land resources and data regarding their change. As an important artificial landscape type, the utilization expansion, abandonment, and conversion of cropland with other types will change both the material flow and energy flow of the cultivated land, affecting the Earth system’s climate, hydrology, biological cycle, and other global change processes [8,9,10,11]. Therefore, in the context of the continuously increasing world population, the decrease in per capita cropland, and the great challenge of food security, it is of great significance to quickly grasp the quantity, spatial distribution, and spatiotemporal characteristics of the existing cropland resources for the study of sustainable development, global environmental change, and strategic decisions regarding food security internationalization [12,13,14,15].
With the continuous development of Earth observation technology, remote sensing has become an important means for obtaining the spatial distribution information of land cover types, including cropland [16,17,18,19,20,21]. At present, a variety of land cover datasets that include cropland types have emerged [22,23], such as that of the University of Maryland [24], the Global Land Cover of the European Space Agency (ESA) [25], and the GLC2000 produced by the Joint Research Centre of the European Commission [26]. These data are widely used in global food security decision-making as well as the analysis of changes in agricultural land systems. Even with the continuous progress of global research on climate change prediction, biodiversity conservation, and sustainable development, new requirements are gradually being put forward for global cropland monitoring data with high spatiotemporal resolution. These requirements have led to datasets such as the Finer Resolution Observation and Monitoring of Global Land Cover (FROM-GLC) [27] produced by Tsinghua University, the GlobeLand30 [28] produced by the National Geomatics Center of China, the Global Food Security-support Analysis Data at 30 m for Southeast Asia (GFSAD30SEACE) [29] led by the United States Geological Survey (USGS), and the Global Land Cover Fine Surface Cover (GLC_FCS30-2015) [30] produced by the Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences. The higher spatial resolution of these products makes them more suitable for depicting the land use activities in fragmented areas of the surface landscape than similar products and affords them a unique advantage when describing changes in global cropland spatiotemporal patterns. Though it cannot be ignored, the differences in remote sensing data, spatial resolution, classification systems, and mapping techniques inevitably lead to differences among regional or global surface coverage datasets. Therefore, when conducting comparative analyses and identification, the similarities and differences among these land cover products are of great significance to data producers and users [31,32].
At present, scientists in China and abroad have carried out evaluation and consistency research on multi-source remote sensing land cover data. For example, Liang et al. [33] studied the accuracy and consistency of four global land cover datasets in the Arctic, discovering that the overall accuracy (OA) of the CCI-LC2000 dataset was the highest, with a value of 63.5%, while that of the MODIS dataset was the lowest, only 29.5%. Xu et al. [34] analyzed the three datasets covering Africa in terms of validation samples and statistics from the Food and Agriculture Organization (FAO) of the United Nations. The accuracy of these datasets was >60%, with the CGLS-LC100 data being the most consistent with the FAO statistics. The consistency analysis of the MODIS and GLC2000 datasets by Giri et al. [35] revealed that their consistency for land cover types is generally better, with the exception of tropical savanna/shrub and wetland, although they exhibited less consistency for finer land cover types. Hua et al. [36] demonstrated that the overall consistency of these datasets ranged from 49.2 to 67.63% according to a consistency study of 5 global land cover datasets covering different climatic and elevation zones. The aforementioned analyses of different land cover data are of great significance for research in the fields of environmental change, land resource surveying, and ecosystem circulation. Thus far, however, evaluation and consistency research of different cropland datasets have been very limited. Vancutsem et al. [37] fused 10 land cover datasets covering the African region in order to obtain cropland data with a resolution of 250 m, subsequently evaluating and analyzing their accuracy in terms of 2 cropland datasets with a spatial resolution of 1 km. Their results showed that the accuracy of the fused cropland dataset was higher than that of the other two cropland datasets. The research of Vancutsem uncovered the fact that more accurate cropland data can be obtained through a spatial combination of multiple existing cropland datasets. Chen et al. [38] studied the MODIS2010, GlobCover2009, FROM-GLC2010, and GlobeLand30-2010 global land cover datasets containing cropland types. Their results revealed that the OA of these datasets for cropland ranged from 61.26 to 80.63%, with the accuracy of the GlobeLand30-2010 dataset being the highest. These evaluations and consistency studies on cropland datasets provide important reference information for research in areas such as food security assessment.
The existing research, however, has primarily focused on the evaluation and consistency of first-class cropland and has only analyzed the causes of inconsistency in terms of cartography (such as Chen’s research) [38,39]. These studies have neglected the importance of fine cropland types (e.g., the paddy) for topics such as global food security assessment and failed to comprehensively explore the causes of inconsistency from the perspectives of both cartography and geography (e.g., elevation). Moreover, the existing research validation samples have mainly been obtained based on Google Earth, without the validation of field survey data, thus indicating that their credibility has room for improvement [40,41]. At the same time, the data sources targeted by existing research have tended to be those with low spatiotemporal resolution [42,43]. Cambodia, as a tropical country focused on agriculture, is rich in soil and water and has a climate suitable for year-round crop growth. Agriculture is the first pillar of the industry in Cambodia’s economy, and agricultural residents account for 85% of its total population, indicating that this country has great potential to expand its agricultural production [44]. Cambodia is a poor country with high vegetation coverage and mainly agricultural development. Therefore, the spatial pattern distribution and utilization of land cover in this region deserve attention, so as to provide reference for the future policy formulation of land use planning and rational utilization of land resources by the Cambodian government.
Hence, in order to make up for the shortcomings of existing research, this investigation took Cambodia as a case study and utilized areal, spatial distribution pattern, elevation classification, and field survey data assessment methods to analyze the GFSAD30SEACE, GLC_FCS30-2015, FROM_GLC2015, and SERVIR-Mekong cropland datasets and comprehensively examined the causes of inconsistency in terms of geography and cartography. The results not only provide reference for the selection of appropriate land cover data in the study of global climate change, the analysis of sustainable cropland resources, agricultural monitoring, and food security assessment but also provide information for countries attempting to carry out transnational cropland investment and global food security strategy decision-making.
2 Study area and data
2.1 Study area
Cambodia is located on the Indochina Peninsula, with geographic coordinates of 10°21′–14°32′N and 102°18′–107°37′E. Cambodia borders Thailand to the west and northwest, Laos to the northeast, Vietnam to the east and southeast, and the Gulf of Thailand to the south. Major water bodies include the Mekong River and Southeast Asia’s largest freshwater lake, Tonle Sap Lake. Cambodia is a traditional agricultural country, of which agricultural residents account for 85% of its total population and 78% of the national labor force [45]. Cambodia’s main agricultural crop is rice, and there are also many special economic crops, such as cassava, sugar cane, tropical fruits, and so on. In addition, the rubber industry is an important part of Cambodia’s agricultural income. Altogether, the agricultural output value accounts for approximately 30% of Cambodia’s GDP [46]. The terrain of Cambodia is a dish-shaped basin. The central and southern sections are plains, and the eastern, northern, and western regions are surrounded by plateaus and mountains, forming a landform with high edges, a low central area, and an opening to the southeast. Cambodia has a typical tropical monsoon climate, with an average annual temperature of 29–30°C, a rainy season from May to October, and a dry season from November to April [47]. Under the influence of topography and the East Asian monsoon, precipitation varies greatly from region to region, with annual average rainfall of 1,250–1,750 mm [47]. Figure 1 depicts the location and topography of the study area.
Location and topographic map of Cambodia.
2.2 Data and preprocessing
In order to reduce the error caused by differences in temporal and spatial resolutions, this study selected four datasets from the current high-resolution land cover products for analysis, namely, the GFSAD30SEACE (https://lpdaac.usgs.gov/products/gfsad30seacev001/), GLC_FCS30-2015 (http://data.casearth.cn/sdo/detail/5d904b7a0887164a5c7fbfa0), FROM_GLC2015 (http://data.ess.tsinghua.edu.cn/), and SERVIR-Mekong (https://rlcms-servir.adpc.net/en/landcover/) led by the USAID, NASA, and other teams. The main parameters of these datasets are listed in Table 1.
Parameters of the four land cover datasets
| Dataset | Resolution (m) | Year | Classification number | Classification method | Publication organization | Data source/sensor | OA (%) | Coverage |
|---|---|---|---|---|---|---|---|---|
| GFSAD30SEACE | 30 | 2015 | 3 | Random forest | USGS, University of New Hampshire, University of Wisconsin- Madison, and others | Landsat ETM+/OLI | 88.6 | Global |
| GLC_FCS30-2015 | 30 | 2015 | 30 | Random forest | Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences | Landsat OLI | 81.4 | Global |
| FROM_GLC2015 | 30 | 2015 | 26 | Random forest | Tsinghua University | Landsat TM/ETM+/OLI | 77.3 | Global |
| SERVIR-Mekong | 30 | 2015 | 22 | Support vector machine and Random forest | USAID, NASA, and ADPC | |||
| Landsat TM/ETM+ | Not published | Indochina peninsula |
When performing consistency analysis, the original land cover datasets were required to undergo a series of preprocessing steps. First, based on ArcGIS software, the cropland datasets in the study area were obtained through the Cambodian vector boundary, and the coordinate systems were transformed into the Albers projection, in the WGS84 coordinate system. Additionally, in order to unify the classification systems of the four datasets, this study employed the definition of cropland in the GFSAD30SEACE dataset as a reference, standardized the same code types in the original classification systems of the other land cover datasets, and established a unified cropland classification system. The correspondences between the normalized classification system and the original classification systems after data preprocessing are listed in Table 2.
Correspondences between the normalized classification system and the original classification systems included in the study area
| Normalized type | GFSAD30SEACE | GLC_FCS30-2015 | FROM_GLC2015 | SERVIR-Mekong | |
|---|---|---|---|---|---|
| Cropland | Cropland | ||||
| Paddy | — | Irrigated cropland | Rice paddy | Rice paddy | |
| Orchard | — | Tree or shrub cover (orchard) | Orchard | — | |
| Other cropland | — | Rainfed cropland/herbaceous cover | Other cropland/bare farmland | Cropland | |
| Non-cropland | Water/non-cropland | Others | Others | Others | |
Note: “—” indicates no data.
For the accuracy verification of the cropland areas from the four datasets, we introduced the FAO (http://www.fao.org/faostat/zh/#search/cropland) statistical data. The FAO is an international organization established specifically for agriculture and is expressly focused on agricultural data statistics. The FAO statistical agency conducts an annual census, and its statistical data are collected in a variety of ways, mainly involving the regular sending of feedback information obtained from agricultural survey questionnaires in order to determine statistical data; gathering statistical data from national statistical yearbooks, public publications, and official national statistical websites; collecting relevant information through remote sensing and geographic information system technologies; and using remote sensing and geographic information system technologies and field surveys during the harvest seasons of key crops [48]. According to FAO statistics, the cropland land cover type is defined as the land used to grow crops, including arable land and permanent crops. Of these, arable land includes temporary cropland for arable land use only, temporary grassland for mowing or grazing, vegetable garden land, and temporarily fallow land (<5 years). Abandoned land resulting from the transfer of cropland is not included in arable land. Permanent crops are the land areas cultivated with long-term crops that do not have to be replanted for several years (such as cocoa and coffee), as well as the land under trees and shrubs producing flowers (except forest trees) and nurseries. Permanent grasslands and grazing land are not included in the permanent crops. In addition, in FAO statistics, crop area is the surface area of the land on which a crop is grown [49].
3 Methods
3.1 Spatial pattern consistency
Based on ArcGIS software, we used the spatial superposition method to calculate the consistency of land cover data at each grid point by pixel. For example, when analyzing the spatial consistency map of cropland types, the value of a given cropland type in the 4 datasets was set to 1, and those of the other types were set to 0. Next the pixel-based superposition calculation of the four cropland datasets was performed using the grid calculator. In the final output, if the pixel value on a grid point was four, this meant that all four datasets defined that grid point as cropland. The same number was defined according to the type of cropland that was counted by each pixel, and the degree of consistency was divided into four grades: (1) complete inconsistency: all four data indication types were different; (2) low consistency: two data indication types were the same; (3) high consistency: three data indication types were the same; (4) complete consistency: all four data indication types were the same.
Topographic features exert an impact on the accuracy of land cover data [50]. This study introduced elevation information in order to analyze the consistency distribution rule of multi-source remote sensing cropland data for different elevations. Based on the topographic and geomorphic characteristics of the study area, in order to provide the land cover types with obvious gradient characteristics in the elevation distribution, we divided the elevation into five grades (Table 3) [51].
Elevation classification
| Grade | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| Elevation (m) | <50 | 50–100 | 100–200 | 200–300 | >300 |
3.2 Precision evaluation based on field survey data
The confusion matrix is an accuracy evaluation method that uses the form of a matrix to calculate the consistency between the classification result of samples and the actual category of a surface [52,53]. The precision indices obtained from the confusion matrix include (1) OA, which provides users with the probability that randomly selected locations in the classification results will be correctly classified; (2) user accuracy (UA), which provides users with the conditional probability that the actual classification is correct in a specific region; (3) producer accuracy (PA), which represents the conditional probability that the ground truth category is correctly classified; and (4) the kappa coefficient, which is often used as an overall measure of accuracy. It is calculated by multiplying the total number of pixels in all the surface truth classifications by the sum of the diagonal lines of the confusion matrix, then subtracting the product of the total number of ground truth pixels of a certain class and the total number of classified pixels of the class and summing the result of all categories, and then dividing by the square of the total number of pixels minus the product of the total number of surface truth pixels of a certain class and the total number of classified pixels of the class. The calculation formulas of these indicators are as follows [54,55]:
where x ii is the correctly classified pixel number of type i; n is the total pixel number in the study area; x i+ is the total pixel number of type i in the dataset to be verified; x +i is the total pixel number of type i in the reference data; and r represents the number of rows in the confusion matrix.
In this study, absolute accuracy evaluation was carried out using the field survey samples in the study area in 2018 and combined with the samples obtained by Google Earth images. The field survey sampling process consisted of five steps: point selection, coordinate location, photo shooting, recording, and correction. The sampling process randomly selected field survey samples according to the composition of land cover types. At the same time, given that the base year of the data to be evaluated was 2015, there was a time lag before the field survey took place. Therefore, the Google Earth images of the 2015 timeline were used to check the survey samples in order to reduce the evaluation error caused by sample quality. Through the above steps, a total of 216 field study samples were obtained. Additionally, in areas with complex terrain, the field survey sampling is limited by transportation challenges and cost. In contrast, the Google Earth image repository is one of the main data sources for accuracy evaluation due to its advantages of accurate location, high resolution, and wide coverage [56]. Hence, we combined Google Earth image samples to obtain the final validation samples. In order to reduce the adverse effects of positioning and interpretation error on sample quality, the following principles were followed during sample selection and interpretation: (1) Since the localization error of Google Earth image samples is approximately 15 m, the spatial resolution of the four datasets is 30 m. In order to reduce the influence of localization error, the sample point selection area size was 7 × 7 pixels, i.e., 210 m × 210 m from the center point of the homogeneous area. (2) In order to reduce the interpretation error caused by the data time, the 2015 remote sensing images were used, consistent with the time of the data to be evaluated. (3) For some samples that were difficult to interpret, other information was combined to assist interpretation, such as Geo-Wiki [57]. (4) A multi-person independent interpretation was adopted, and a sample was abandoned if the interpretation result could not be unified after negotiation. Based on the above principles, 792 validation samples covering the study area were obtained (Table 4 and Figure 2).
Number of validation samples and photos of field survey
| Type | Sample number (Google Earth) | Sample number (field survey) | Field survey photos |
|---|---|---|---|
| Paddy | 156 | 100 |
|
| Orchard | 65 | 36 |
|
| Other cropland | 57 | 22 |
|
| Non-cropland | 298 | 58 |
|
| Total | 576 | 216 |
Spatial distribution of verification samples.
4 Results
4.1 Comparisons of cropland areas
4.1.1 Comparisons of cropland areas among different land cover datasets
Figure 3 presents the results of the first-class cropland areas from three datasets. The consistency of the cropland extraction areas between the GFSAD30SEACE, GLC_FCS30-2015, and SERVIR-Mekong datasets was high, and the difference of the cropland extraction area as a percentage of the total study area was <0.35%. The difference between the percentage of cropland area from the FROM_GLC2015 data and the other 3 datasets was <2%. Therefore, the analysis of the area statistics revealed that the consistency of the first-class cropland area among the four datasets was high.
Area comparison of level 1 class cropland among four datasets.
The statistical analysis of the areas of the fine class of cropland from the GLC_FCS30-2015, FROM_GLC2015, and SERVIR-Mekong datasets (Figure 4) revealed that the areas of paddy and orchard types extracted from the four datasets varied greatly, and hence the consistency was low. The consistency of three of the datasets was high for other cropland types, and the percentages of their area to the total study area were 34.35% (GLC_FCS30-2015), 38.91% (FROM_GLC2015), and 26.78% (SERVIR-Mekong). These results indicate that although the consistency of the extraction area of the first-class cropland was high among the different datasets, the consistency of the fine second-class cropland was significantly reduced.
Area comparison of level 2 class cropland among three datasets: (a) GLC_FCS30-2015, (b) FROM_GLC2015, and (c) SERVIR-Mekong.
4.1.2 Comparisons between classification results and cropland area statistics
The comparative analysis between the extraction areas of the four cropland datasets and the FAO statistical results showed that for the first-class cropland type (Table 5), the consistency was low, and the error between the cropland extraction areas of the four datasets and the statistical results was >81%. Table 6 presents the comparison between paddy areas of the GLC_FCS30-2015, FROM_GLC2015, and SERVIR-Mekong datasets and that of the FAO statistics. The paddy area extracted from the FROM_GLC2015 dataset exhibited the lowest consistency with the FAO statistical results, with the relative error between them reaching as high as 95.58%. In contrast, the consistency between the paddy area extracted from the SERVIR-Mekong dataset and the FAO statistical results was the highest, with a relative error of only 0.2% between them.
Comparison of the extracted and statistical areas of level 1 class cropland from the four datasets
| Product | GFSAD30SEACE | GLC_FCS30-2015 | FROM_GLC2015 | SERVIR-Mekong |
|---|---|---|---|---|
| RSEA (km2) | 76873.5 | 76228.1 | 73829.6 | 76529.1 |
| SA (km2) | 40660.0 | |||
| REA (%) | 89.06 | 87.48 | 81.58 | 88.22 |
Note: RSEA represents the area extracted by remote sensing, SA represents the statistical area, and REA represents the relative error (REA = ((SA −RSEA)/SA)*100).
Comparison of the extracted and statistical areas of the paddy land cover type from three datasets
| Product | GLC_FCS30-2015 | FROM_GLC2015 | SERVIR-Mekong |
|---|---|---|---|
| RSEA (km2) | 12728.2 | 1237.8 | 28039.7 |
| SA (km2) | 27984.5 | ||
| REA (%) | 54.52 | 95.58 | 0.20 |
Note: RSEA represents the area extracted by remote sensing, SA represents the statistical area, and REA represents the relative error (REA = ((SA −RSEA)/SA)*100).
4.2 Accuracy evaluation based on field survey
Calculation of the confusion matrices for the four cropland datasets based on the field survey data (Tables 7–10) revealed that the OA and kappa coefficient of the cropland from the GFSAD30SEACE dataset were the highest, with values of 87.25% and 0.74, respectively, and the mapping accuracy of the cropland was as high as 88.99%. The OA of the GLC_FCS30-2015 and SERVIR-Mekong datasets was slightly lower than that of the GFSAD30SEACE dataset: 82.20 and 82.83%, respectively. Meanwhile, the OA and kappa coefficient of the cropland from the FROM_GLC2015 dataset were the lowest, with values of 72.35% and 0.45, respectively.
Level 1 class cropland confusion matrix of the GFSAD30SEACE dataset
| Non-cropland | Cropland | Row total | |
|---|---|---|---|
| Non-cropland | 303 | 48 | 351 |
| Cropland | 53 | 388 | 441 |
| Column total | 356 | 436 | 792 |
| PA (%) | 85.11 | 88.99 | |
| UA (%) | 86.32 | 87.98 | |
| OA (%) | 87.25 | ||
| Kappa | 0.74 | ||
Level 1 class cropland confusion matrix of the GLC_FCS30-2015 dataset
| Non-cropland | Cropland | Row total | |
|---|---|---|---|
| Non-cropland | 272 | 57 | 329 |
| Cropland | 84 | 379 | 463 |
| Column total | 356 | 436 | 792 |
| PA (%) | 76.40 | 86.93 | |
| UA (%) | 82.67 | 81.86 | |
| OA (%) | 82.20 | ||
| Kappa | 0.64 | ||
Level 1 class cropland confusion matrix of the FROM_GLC2015 dataset
| Non-cropland | Cropland | Row total | |
|---|---|---|---|
| Non-cropland | 278 | 141 | 419 |
| Cropland | 78 | 295 | 373 |
| Column total | 356 | 436 | 792 |
| PA (%) | 78.09 | 67.66 | |
| UA (%) | 66.35 | 79.09 | |
| OA (%) | 72.35 | ||
| Kappa | 0.45 | ||
Level 1 class cropland confusion matrix of the SERVIR-Mekong dataset
| Non-cropland | Cropland | Row total | |
|---|---|---|---|
| Non-cropland | 288 | 67 | 355 |
| Cropland | 69 | 368 | 437 |
| Column total | 357 | 435 | 792 |
| PA (%) | 80.67 | 84.60 | |
| UA (%) | 81.13 | 84.21 | |
| OA (%) | 82.83 | ||
| Kappa | 0.65 | ||
Similarly, we used field survey data to calculate the confusion matrices of the fine cropland types from the GLC_FCS30-2015, FROM_GLC2015, and SERVIR-Mekong datasets (Tables 11–13). The results revealed that the OA levels of the fine cropland types of the GLC_FCS30-2015 and FROM_GLC2015 datasets were <50%. Meanwhile, the OA and kappa coefficient of the fine cropland types from the SERVIR-Mekong dataset were the highest, with values of 74.87% and 0.61, respectively. The mapping accuracy of the GLC_FCS30-2015 and FROM_GLC2015 datasets was <18% for the paddy and orchard types, while that of the paddy type from the SERVIR-Mekong dataset was 77.73%. The absolute precision evaluation experiment showed that the OA of the first-class cropland from different land cover datasets was higher, although the OA of the more refined cropland types was significantly reduced, especially for the GLC_FCS30-2015 and FROM_GLC2015 datasets.
Level 2 class cropland confusion matrix of the GLC_FCS30-2015 dataset
| Non-cropland | Paddy | Orchard | Other cropland | Row total | |
|---|---|---|---|---|---|
| Non-cropland | 272 | 28 | 18 | 11 | 329 |
| Paddy | 16 | 44 | 6 | 5 | 71 |
| Orchard | 0 | 0 | 13 | 0 | 13 |
| Other cropland | 68 | 184 | 64 | 63 | 379 |
| Column total | 356 | 256 | 101 | 79 | 792 |
| PA (%) | 76.40 | 17.19 | 12.87 | 79.75 | |
| UA (%) | 82.67 | 61.97 | 100.00 | 16.62 | |
| OA (%) | 49.49 | ||||
| Kappa | 0.31 | ||||
Level 2 class cropland confusion matrix of the FROM_GLC2015 dataset
| Non-cropland | Paddy | Orchard | Other cropland | Row total | |
|---|---|---|---|---|---|
| Non-cropland | 278 | 43 | 61 | 37 | 419 |
| Paddy | 0 | 2 | 0 | 0 | 2 |
| Orchard | 2 | 0 | 2 | 0 | 4 |
| Other cropland | 76 | 211 | 38 | 42 | 367 |
| Column total | 356 | 256 | 101 | 79 | 792 |
| PA (%) | 78.09 | 0.78 | 1.98 | 53.16 | |
| UA (%) | 66.35 | 100.00 | 50.00 | 11.44 | |
| OA (%) | 40.91 | ||||
| Kappa | 0.17 | ||||
Level 2 class cropland confusion matrix of the SERVIR-Mekong dataset
| Non-cropland | Paddy | Other cropland | Row total | |
|---|---|---|---|---|
| Non-cropland | 288 | 1 | 67 | 356 |
| Paddy | 7 | 199 | 7 | 213 |
| Other cropland | 61 | 56 | 106 | 223 |
| Column total | 356 | 256 | 180 | 792 |
| PA (%) | 80.90 | 77.73 | 58.89 | |
| UA (%) | 80.90 | 93.43 | 47.53 | |
| OA (%) | 74.87 | |||
| Kappa | 0.61 | |||
4.3 Analysis of spatial pattern consistency
Figure 5 shows the spatial pattern distribution of first-class cropland from the four datasets. The cropland of Cambodia is mainly distributed around Tonle Sap Lake and the central and southern plains. The eastern, northern, and western parts of Cambodia have less cropland due to the plateau and mountainous surroundings, especially the Elephant Mountains in the southwest and the Cardamom Mountains in the west. The analysis of the spatial distribution pattern revealed that the GFSAD30SEACE, GLC_FCS30-2015, and SERVIR-Mekong datasets exhibited high overall consistency. The cropland distribution of the GFSAD30SEACE dataset was less than that of either the GLC_FCS30-2015 or SERVIR-Mekong in the marginal area of Tonle Sap Lake, and the spatial patterns in the region varied greatly. The spatial distribution of cropland from the FROM_GLC2015 dataset was more widespread than the distributions of the other three datasets in Kratie and Mondulkiri provinces in eastern Cambodia and Ratanakiri and Stung Treng provinces in northeastern Cambodia, and was quite different from the other three datasets.
Spatial distribution patterns of level 1 class cropland from four different land cover datasets: (a) GFSAD30SEACE, (b) GLC_FCS30-2015, (c) FROM_GLC2015, and (d) SERVIR-Mekong.
The analysis of the spatial distribution pattern of fine cropland (Figure 6) showed that the paddy type from the SERVIR-Mekong dataset was mainly distributed along the Mekong River, Tonle Sap Lake, and the area around the Tonle Sap River. The paddy type from the GLC_FCS30-2015 dataset was less distributed around Tonle Sap Lake and more along the Mekong River and Tonle Sap River. Meanwhile, there was very little paddy type extracted from the FROM_GLC2015 dataset, scattered along the Mekong River, Tonle Sap Lake, and the Tonle Sap River. These results indicate that the spatial patterns of the paddy land cover type varied significantly among these three datasets. For orchard types, the GLC_FCS30-2015 dataset showed orchards concentrated in Kampong Cham, Kampong Thom, and Tbong Khmum provinces in the plain region of southeast Cambodia, while orchards extracted from the FROM_GLC2015 dataset were scattered in the eastern, southern, and northern regions of the country. This indicates that the spatial distribution patterns of orchard types from these two datasets were quite different.
Spatial distribution patterns of level 2 class cropland from three of the land cover datasets: (a) GLC_FCS30-2015, (b) FROM_GLC2015, and (c) SERVIR-Mekong.
The spatial consistency distribution patterns of different datasets were obtained by the spatial superposition of cropland extracted from four datasets. Figure 7 shows that the spatial consistency of the cropland areas from the four datasets was low; the completely consistent area accounted for 22.53% of the total study area and was mainly distributed in the Tonle Sap Lake plain region in central Cambodia and the Tonle Sap River and Mekong plain sections of southern Cambodia. The completely inconsistent area of the four datasets accounted for 14.13% of the total study area and was concentrated in the mountainous sections of Kratie Province and Mondulkiri Province in eastern Cambodia, as well as Ratanakiri Province and Stung Treng Province in northeast Cambodia, while being scattered in the mountainous regions of northern Cambodia, the central plain, and the narrow coastal plain in the western part of the country. The results presented in Figure 8 reveal that the spatial consistency of the fine cropland class from the four datasets was very low, especially for the paddy and orchard types. The complete consistency area of the paddy type from the GLC_FCS30-2015, FROM_GLC2015, and SERVIR-Mekong datasets only accounted for 0.23% of the total study area, mainly distributed along the Mekong River and Tonle Sap River in the southern part of Cambodia, while the complete consistency area of the orchard type between the GLC_FCS30-2015 and FROM_GLC2015 datasets accounted for only 0.01% of the total study area.
Spatial consistency distribution of level 1 class cropland among the four land cover datasets.
Spatial consistency distribution of level 2 class cropland among the four land cover datasets: (a) paddy, (b) orchard, and (c) other cropland.
4.4 Variation in spatial consistency distribution with elevation
Figures 9 and 10 show the change in the spatial consistency with elevation for the cropland areas from the four datasets. There were many highly consistent and completely consistent areas in the plain regions with elevations <50 m, indicating that the four datasets have better consistency in these areas and that the spatial distribution of these areas was mainly in the central and southern regions of Cambodia. The extent of cropland in these areas is vast and the distribution is more concentrated, and different datasets can accurately extract the range of cropland. When the elevation is >50 m, the spatial consistency of different datasets decreased gradually with increase in elevation. The completely inconsistent areas were mainly concentrated in the elevation ranges of 50–100 and 100–200 m, where the distribution of cropland is relatively fragmented and extraction is difficult.
Consistency distributions at different elevations: (a) complete inconsistency, (b) low consistency, (c) high consistency, and (d) complete consistency.
Consistency distribution areas at different elevations.
5 Discussion
Cropland, as an important surface type in global land cover mapping, plays a significant role in global food security, ecological environment protection, and global climate change. In this study, the accuracy evaluation and consistency analysis of four cropland cover datasets revealed that the cropland types differ among these datasets in terms of area, spatial distribution pattern, and accuracy. The main reasons for these differences include the following two factors.
Geographic factors. The spatial pattern consistency among the different cropland datasets was found to be low, especially for the paddy and orchard types. These areas of inconsistency were mainly distributed in the eastern and northern regions of Cambodia. According to the statistics of the land cover types from different datasets in inconsistent areas at different elevations (Figure 11), the main causes of the inconsistencies stemmed from two determinants: first, in areas with elevations <50 m, land cover types are more complex, the heterogeneity of the surface landscape is higher, and these areas are more affected by human activities; second, in areas with elevations ranging from 50 to 200 m, the spectral characteristics of cropland are similar to those of forests, grassland, and shrubs, making it difficult to accurately distinguish them using remote sensing classification, which leads to the obvious confusion of these types. Therefore, in order to further improve the classification accuracy of cropland types in the future, land cover data producers should adopt the classification strategy of zonal stratification and focus on the classification methods with larger inconsistent areas and more easily confused land cover types.
Mapping technology factors. Remote sensing technology is playing an increasingly significant role in land use information acquisition and has become an important means of thematic information extraction and mapping due to its rapid, accurate, and dynamic characteristics. The basic process of obtaining regional or global land cover products through remote sensing methods consists of remote sensing data acquisition, data preprocessing, multi-feature extraction, model training classification, and accuracy evaluation (Figure 12).
Area statistics of land cover types from three products in inconsistent areas at different elevations: (a) GLC_FCS30-2015, (b) FROM_GLC2015, and (c) SERVIR-Mekong.
Land cover mapping flowchart.
It cannot be ignored, however, that different products may use different satellite images (such as Landsat, MODIS, and GF), classification techniques, classification systems, and so on in the process of remote sensing mapping, and the resulting land cover products themselves are different. Therefore, the reasons for the inconsistency of the four cropland datasets in this study were discussed and analyzed in terms of remote sensing mapping technology, including the following aspects.
(1) The effect of scale size. First-class cropland types with high precision can be obtained based on medium-resolution Landsat images. For second-class cropland types, the OA of the regional-scale SERVIR-Mekong dataset was comparatively high, with a relative error between the paddy extraction area and the FAO statistical area of only 0.2%, while the second-class cropland mapping accuracy of the global scale GLC_FCS30-2015 and FROM_GLC2015 datasets were relatively low. This indicates that the spatial scale of the data to be produced will also affect the final mapping accuracy. This also tells future land cover data producers that using only Landsat remote sensing image to produce second-class global scale land cover data products can no longer meet the accuracy requirements. It is therefore necessary to use higher-resolution images such as SAR, Sentinel, and LiDAR as either auxiliary data or original classification data in order to improve the mapping accuracy of second-class land cover types [58,59,60,61], especially in areas where high-quality remote sensing images are difficult to obtain due to complex weather conditions, such as the frequent rain and clouds found in Southeast Asian countries.
(2) The effect of classification system. Different research purposes may lead to differences in classification systems among various land cover products [31]. The definition of cropland used in this study was the same as the definition in the GFSAD30SEACE dataset, i.e., it included all cropland crops (such as rice, corn, soybeans, cotton, wheat, orchards, coffee, tea, rubber, and palm), as well as agricultural fallow land. In addition to the cropland types defined above, the cropland definition in the GLC_FCS30-2015 dataset included herb-covered cropland, the FROM_GLC2015 dataset included the greenhouse land cover type, while the specific definition of cropland in the SERVIR-Mekong dataset was not clearly delineated. Therefore, differences in the definition of cropland among these different datasets can also have an impact on consistency evaluation. Moreover, the definition of cropland as the land used for planting crops in the FAO statistics, including arable land and permanent crops, is different from the definition of cropland in the four datasets, which also leads to the large difference between the extraction areas of the first-class cropland from the four datasets and the FAO statistical results. Therefore, it is necessary to provide a clear definition of each land cover type in the classification system in order to reduce the series of errors caused by fuzzy concepts during the creation of the classification system used in the subsequent production of land cover data.
(3) The effect of classification strategies and methods. Although the random forest classification algorithm was used in all four land cover datasets, there were differences in the overall process of data production (such as sample acquisition). After the GFSAD30SEACE dataset had combined and preprocessed the acquired time series Landsat images, they were then compared with sub-meter to 5 m very high spatial resolution imagery (VHRI) in order to ensure that no artifacts were introduced into the composites [62]. The GFSAD30SEACE dataset has high credibility in terms of the samples used for classification and verification. It was mainly obtained from a variety of data, including the sub-meter to 5 m VHRI data provided by the National Geospatial Agency, the author’s own field survey data [63], and field survey data provided by tertiary collaborators [64]. In addition, the GFSAD30SEACE dataset used OpenStreetMap vector data downloaded from https://extract.bbbike.org to perform post-processing on the classification results. The GLC_FCS30-2015 dataset processing method was more complex than those of the other three datasets consisting of topographic radiation C correction, atmospheric correction combining MODIS atmospheric products with 6 S radiation transmission models, and cloud and shadow detection. The GLC_FCS30-2015 dataset constructed a global spatiotemporal spectral library (GSPECLib) through the time series MCD43A4 [65] reflectance products and CCI_LC2015 surface cover products [66]. Then, based on the time series Landsat and auxiliary terrain data integrated with the Google Earth Engine, along with the geographic location information of the GSPECLib spectra, multi-temporal classification model training was carried out region by region, finally obtaining the global 30 m fine surface coverage data. Based on the classification strategies and methods used in the GLC_FCS30-2015 dataset, the spatial distribution pattern of ground objects was continuous, and the banding problem in the results of single-time image classification was effectively eliminated. For the FROM_GLC2015 dataset, the authors used Global Mapper software developed by their own team to process the Landsat images. The Earth was first divided into 16 regions, and the training classification was then performed with the first global multi-season sample data [67] scene-by-scene during the classification process, and the classification results were pieced together scene-by-scene. Moreover, the FROM_GLC2015 dataset introduced nighttime light data in order to improve the accuracy of artificial surface types. We found, however, that the banding problem of land cover data obtained by the strategies and methods used in the FROM_GLC2015 dataset was more serious. For the SERVIR-Mekong dataset, field survey information and high-resolution satellite images were combined to obtain sample points, after which Landsat data, surface reflectance composites, Shuttle Radar Topography Mission data, and JRC Global Surface Water Mapping Layers were utilized as variables for model training classification. Through experimentation, we discovered that the accuracy values of the adopted first-class cropland classification strategies and methods used in the GFSAD30SEACE, GLC_FCS30-2015, and SERVIR-Mekong datasets were >82%, indicating that land cover users can utilize these outputs directly. For more refined cropland types, however, only the accuracy of the SERVIR-Mekong dataset was sufficient, with a value of 74.87%, while the accuracy levels of the GLC_FCS30-2015 and FROM_GLC2015 datasets were <50%, and thus cannot meet the requirements of relevant research. Therefore, the classification strategies and methods of the GFSAD30SEACE, GLC_FCS30-2015, and SERVIR-Mekong datasets can be adopted in the subsequent mapping of regional or global scales for first-class cropland. However, only the SERVIR-Mekong dataset can be used as auxiliary data, and it is necessary to integrate the advantages of various data classification schemes for the classification of fine cropland types.
(4) Other factors. Projection conversion and merging preprocessing of classification systems may be external factors affecting the consistency of different datasets. The preprocessing method may lead to the loss of original information in the land cover data or the confusion of information, which will affect the consistency between different datasets. In addition, the error caused by manual visual interpretation cannot be ignored when selecting verification samples involved in accuracy evaluation. The possible errors caused by visual interpretation in this study mainly lie in the size of the selected sample points and the understanding degree of the interpreters on the ground objects in the study area. Of course, since the samples this study selected were finally formed by independent interpretation by many people and combining with field survey data, their reliability was relatively high, and the impact on the research results of this study was minimal.
6 Conclusion
This study analyzed the consistency of four cropland datasets and explored the reasons for the inconsistencies in terms of geography and cartographic technology. The results revealed that although the four datasets exhibited high consistency in the area of first-class cropland, the difference between the cropland definitions of different datasets and that of the FAO data resulted in a large difference between the remotely extracted areas and the FAO statistical results. The accuracy evaluation of the field survey data showed that the OA of the first-class cropland extracted by the classification strategies and methods used in the GFSAD30SEACE, GLC_FCS30-2015, and SERVIR-Mekong datasets was >82%, and hence land cover users can directly employ one of these datasets. For the more refined cropland types, however, due to the influence of scale size, the OA of the regional-scale SERVIR-Mekong dataset was high, i.e., 74.87%, while the OA of the two global-scale datasets was <50%, which does not meet the requirements of the relevant research. Furthermore, in areas with elevations of 50–200 m, it is more difficult for remote sensing techniques to accurately distinguish between cropland and forest, grassland, and shrub types due to their similar spectral characteristics, leading to serious confusion among these types, and thus resulting in high inconsistency among the four datasets in these areas. Therefore, when investigating land cover mapping, we should focus on the classification strategies and methods with high inconsistency and confusing types. Our research results not only provide a reference for land cover users attempting to select suitable cropland data for research on cropland resource monitoring, cropland quality evaluation, and food security assessment but also provide an important reference for land cover producers attempting to obtain land cover datasets quickly and accurately.
Acknowledgments
This study was supported by the National Key Research and Development Program of China, Grant no. 2016YFB0501404; the CAS Earth Big Data Science Project, Grant no. XDA19060303; the National Science Foundation of China, Grant no. 41671436, and the Innovation Project of LREIS, Grant no. O88RAA01YA.
-
Conflict of interest: The authors declare no conflict of interest.
-
Author contributions: Junmei Kang : writing – original draft, methods, formal analysis; Jun Wang: formal analysis, visualization; Mianqing Zhong: software.
References
[1] Deng X, Huang J, Rozelle S, Zhang J, Li Z. Impact of urbanization on cultivated land changes in China. Land Use Policy. 2015;45:1–7.10.1016/j.landusepol.2015.01.007Search in Google Scholar
[2] Lu X, Shi Y, Chen C, Yu M. Monitoring cropland transition and its impact on ecosystem services value in developed regions of China: A case study of Jiangsu Province. Land Use Policy. 2017;69:25–40.10.1016/j.landusepol.2017.08.035Search in Google Scholar
[3] See L, Fritz S, You L, Ramankutty N, Herrero M, Justice C, et al. Improved global cropland data as an essential ingredient for food security. Glob Food Sec. 2015;4:37–45.10.1016/j.gfs.2014.10.004Search in Google Scholar
[4] Foley JA, DeFries R, Asner GP, Barford C, Bonan G, Carpenter SR, et al. Global consequences of land use. Science. 2005;309:570–4.10.1126/science.1111772Search in Google Scholar PubMed
[5] Lambin EF, Meyfroidt P. Global land use change, economic globalization, and the looming land scarcity. Proc Natl Acad Sci U S Am. 2011;108:3465–72.10.1073/pnas.1100480108Search in Google Scholar PubMed PubMed Central
[6] Lambin EF. Conditions for sustainability of human–environment systems: Information, motivation, and capacity. Glob Environ Change. 2005;3:177–80.10.1016/j.gloenvcha.2005.06.002Search in Google Scholar
[7] Yin Y, Pierce JT. Integrated resource assessment and sustainable land use. Environ Manag. 1993;17:319–27.10.1007/BF02394674Search in Google Scholar
[8] Yu Q, Hu Q, van Vliet J, Verburg PH, Wu W. GlobeLand30 shows little cropland area loss but greater fragmentation in China. Int J Appl Earth Observ Geoinf. 2018;66:37–45.10.1016/j.jag.2017.11.002Search in Google Scholar
[9] Cui Y, Liu J, Xu X, Dong J, Li N, Fu Y, et al. Accelerating cities in an unsustainable landscape: urban expansion and cropland occupation in China, 1990–2030. Sustainability. 2019;11:2283.10.3390/su11082283Search in Google Scholar
[10] Li H, Zhang Q, Singh VP, Shi P, Sun P. Hydrological effects of cropland and climatic changes in arid and semi-arid river basins: a case study from the Yellow River basin, China. J Hydrol. 2017;549:547–57.10.1016/j.jhydrol.2017.04.024Search in Google Scholar
[11] Ouyang W, Gao X, Hao Z, Liu H, Shi Y, Hao F. Farmland shift due to climate warming and impacts on temporal-spatial distributions of water resources in a middle-high latitude agricultural watershed. J Hydrol. 2017;547:156–67.10.1016/j.jhydrol.2017.01.050Search in Google Scholar
[12] Wang G, Liu Y, Li Y, Chen Y. Dynamic trends and driving forces of land use intensification of cultivated land in China. J Geograph Sci. 2015;25:45–57.10.1007/s11442-015-1152-4Search in Google Scholar
[13] Yang X, Jin X, Guo B, Long Y, Zhou Y. Research on reconstructing spatial distribution of historical cropland over 300 years in traditional cultivated regions of China. Glob Planet Change. 2015;128:90–102.10.1016/j.gloplacha.2015.02.007Search in Google Scholar
[14] Delzeit R, Zabel F, Meyer C, Václavík T. Addressing future trade-offs between biodiversity and cropland expansion to improve food security. Reg Environ Change. 2017;17:1429–41.10.1007/s10113-016-0927-1Search in Google Scholar
[15] Bharathkumar L, Mohammed-Aslam M. Crop pattern mapping of Tumkur taluk using NDVI technique: a remote sensing and GIS approach. Aquat Proc. 2015;4:1397–404.10.1016/j.aqpro.2015.02.181Search in Google Scholar
[16] Rawat J, Kumar M. Monitoring land use/cover change using remote sensing and GIS techniques: A case study of Hawalbagh block, district Almora, Uttarakhand, India. Egypt J Remote Sens Space Sci. 2015;18:77–84.10.1016/j.ejrs.2015.02.002Search in Google Scholar
[17] Yifang B, Gong P, Gini C. Global land cover mapping using Earth observation satellite data: Recent progresses and challenges. ISPRS J Photogram Remote Sens (Print). 2015;103:1–6.10.1016/j.isprsjprs.2015.01.001Search in Google Scholar
[18] Abdikan S, Sanli FB, Ustuner M, Calò F. Land cover mapping using sentinel-1 SAR data. Int Arch Photogramm Remote Sens Spat Inf Sci. 2016;41:757.10.5194/isprs-archives-XLI-B7-757-2016Search in Google Scholar
[19] Roy PS, Roy A, Joshi PK, Kale MP, Srivastava VK, Srivastava SK, et al. Development of decadal (1985–1995–2005) land use and land cover database for India. Remote Sens. 2015;7:2401–30.10.3390/rs70302401Search in Google Scholar
[20] Yan F, Liu X, Chen J, Yu L, Yang C, Chang L, et al. China’s wetland databases based on remote sensing technology. Chin Geograph Sci. 2017;27:374–88.10.1007/s11769-017-0872-zSearch in Google Scholar
[21] Thenkabail PS, Teluguntla PG, Xiong J, Oliphant A, Congalton RG, Ozdogan M, et al. Global cropland-extent product at 30-m resolution (GCEP30) derived from Landsat satellite time-series data for the year 2015 using multiple machine-learning algorithms on Google Earth engine cloud; 2330-7102. US Geological Survey; 2021.10.3133/pp1868Search in Google Scholar
[22] Arino O, Gross D, Ranera F, Leroy M, Bicheron P, Brockman C, et al. GlobCover: ESA service for global land cover from MERIS. Proceedings of 2007 IEEE International Geoscience and Remote Sensing Symposium. 2008; p. 2412–5.10.1109/IGARSS.2007.4423328Search in Google Scholar
[23] Nowak DJ, Greenfield EJ. Evaluating the national land cover database tree canopy and impervious cover estimates across the conterminous United States: a comparison with photo-interpreted estimates. Environ Manag. 2010;46:378–90.10.1007/s00267-010-9536-9Search in Google Scholar PubMed PubMed Central
[24] Hansen MC, DeFries RS, Townshend JR, Sohlberg R. Global land cover classification at 1 km spatial resolution using a classification tree approach. Int J Remote Sens. 2000;21:1331–64.10.1080/014311600210209Search in Google Scholar
[25] Bontemps S, Defourny P, Van Bogaert E, Arino O, Kalogirou V, Perez JR. GLOBCOVER 2009-Products description Valid Rep. 2011;2. http://ionia1.esrin.esa.int/docs/GLOBCOVER2009_Validation_Report_2.Search in Google Scholar
[26] Bartholome E, Belward AS. GLC2000: a new approach to global land cover mapping from Earth observation data. Int J Remote Sens. 2005;26:1959–77.10.1080/01431160412331291297Search in Google Scholar
[27] Gong P, Wang J, Yu L, Zhao Y, Zhao Y, Liang L, et al. Finer resolution observation and monitoring of global land cover: First mapping results with Landsat TM and ETM+ data. Int J Remote Sens. 2013;34:2607–54.10.1080/01431161.2012.748992Search in Google Scholar
[28] Chen J, Chen J, Liao A, Cao X, Chen L, Chen X, et al. Global land cover mapping at 30 m resolution: A POK-based operational approach. ISPRS J Photogram Remote Sens. 2015;103:7–27.10.1016/j.isprsjprs.2014.09.002Search in Google Scholar
[29] Oliphant AJ, Thenkabail PS, Teluguntla P, Xiong J, Gumma MK, Congalton RG, et al. Mapping cropland extent of Southeast and Northeast Asia using multi-year time-series Landsat 30-m data using a random forest classifier on the Google Earth Engine Cloud. Int J Appl Earth Observ Geoinf. 2019;81:110–24.10.1016/j.jag.2018.11.014Search in Google Scholar
[30] Zhang X, Liu L, Chen X, Xie S, Gao Y. Fine land-cover mapping in China using landsat datacube and an operational SPECLib-based approach. Remote Sens. 2019;11:1056.10.3390/rs11091056Search in Google Scholar
[31] Congalton RG, Gu J, Yadav K, Thenkabail P, Ozdogan M. Global land cover mapping: A review and uncertainty analysis. Remote Sens. 2014;6:12070–93.10.3390/rs61212070Search in Google Scholar
[32] Wang Z, Liu L. Assessment of coarse-resolution land cover products using CASI hyperspectral data in an arid zone in northwestern China. Remote Sens. 2014;6:2864–83.10.3390/rs6042864Search in Google Scholar
[33] Liang L, Liu Q, Liu G, Li H, Huang C. Accuracy evaluation and consistency analysis of four global land cover products in the arctic region. Remote Sens. 2019;11:1396.10.3390/rs11121396Search in Google Scholar
[34] Xu Y, Yu L, Feng D, Peng D, Li C, Huang X, et al. Comparisons of three recent moderate resolution African land cover datasets: CGLS-LC100, ESA-S2-LC20, and FROM-GLC-Africa30. Int J Remote Sens. 2019;40:6185–202.10.1080/01431161.2019.1587207Search in Google Scholar
[35] Giri C, Zhu Z, Reed B. A comparative analysis of the Global Land Cover 2000 and MODIS land cover datasets. Remote Sens Environ. 2005;94:123–32.10.1016/j.rse.2004.09.005Search in Google Scholar
[36] Hua T, Zhao W, Liu Y, Wang S, Yang S. Spatial consistency assessments for global land-cover datasets: A comparison among GLC2000, CCI LC, MCD12, GLOBCOVER and GLCNMO. Remote Sens. 2018;10:1846.10.3390/rs10111846Search in Google Scholar
[37] Vancutsem C, Marinho E, Kayitakire F, See L, Fritz S. Harmonizing and combining existing land cover/land use datasets for cropland area monitoring at the African continental scale. Remote Sens. 2013;5:19–41.10.3390/rs5010019Search in Google Scholar
[38] Chen X-y, Lin Y, Zhang M, Yu L, Li H-C, Bai Y-Q. Assessment of the cropland classifications in four global land cover datasets: A case study of Shaanxi Province, China. J Integr Agric. 2017;16:298–311.10.1016/S2095-3119(16)61442-9Search in Google Scholar
[39] Wang H, Cai L, Wen X, Fan D, Wang Y. Land cover change and multiple remotely sensed datasets consistency in China. Ecosyst Health Sustain. 2022;8:2040385.10.1080/20964129.2022.2040385Search in Google Scholar
[40] See L, Fritz S, Perger C, Schill C, McCallum I, Schepaschenko D, et al. Harnessing the power of volunteers, the internet and Google Earth to collect and validate global spatial information using Geo-Wiki. Technol Forecast Soc Change. 2015;98:324–35.10.1016/j.techfore.2015.03.002Search in Google Scholar
[41] Bey A, Díaz S-P, Maniatis A, Marchi D, Mollicone G, Ricci D, et al. M. Collect earth: Land use and land cover assessment through augmented visual interpretation. Remote Sens. 2016;8:807.10.3390/rs8100807Search in Google Scholar
[42] Wu W, Shibasaki R, Yang P, Ongaro L, Zhou Q, Tang H. Validation and comparison of 1 km global land cover products in China. Int J Remote Sens. 2008;29:3769–85.10.1080/01431160701881897Search in Google Scholar
[43] Herold M, Mayaux P, Woodcock C, Baccini A, Schmullius C. Some challenges in global land cover mapping: An assessment of agreement and accuracy in existing 1 km datasets. Remote Sens Environ. 2008;112:2538–56.10.1016/j.rse.2007.11.013Search in Google Scholar
[44] Li Z. Research on China’s farmland investment in Cambodia. Hubei: Huazhong University of Science & Technology; 2015. p. 1-z.Search in Google Scholar
[45] Gaughan AE, Stevens FR, Linard C, Jia P, Tatem AJ. High resolution population distribution maps for Southeast Asia in 2010 and 2015. PLoS One. 2013;8:e55882.10.1371/journal.pone.0055882Search in Google Scholar PubMed PubMed Central
[46] Lach S, Payanun K, Intaratat K, Sombunsooke B. Roles of agricultural extension policymakers in agricultural development of Cambodia. KASETSART J. 2002;23:167.Search in Google Scholar
[47] Thoeun HC. Observed and projected changes in temperature and rainfall in Cambodia. Weather Clim Extremes. 2015;7:61–71.10.1016/j.wace.2015.02.001Search in Google Scholar
[48] Tang J. Food and agriculture organization of the United Nations (FAO) data quality management system and its revelation to China. Hunnan: Hunnan University; 2013.Search in Google Scholar
[49] Pocketbook FS. World food and agriculture. Rome, Italy: FAO; 2015.Search in Google Scholar
[50] Qi W, Yang X, Li Z, Li Y, Yang F. Correlation of topography and land use type distribution: taking Jinggangshan region in Jiangxi Province for an example. Remote Sens Inf. 2018;33:64–71.Search in Google Scholar
[51] Ma Y, Ji S, Lou S, Liu Z. Spatial-temporal pattern of farmland multiple cropping index in Cambodia. J Hangzhou Norm Univ (Nat Sci Ed). 2014;13(4):418–22.Search in Google Scholar
[52] Canters F. Evaluating the uncertainty of area estimates derived from fuuy land-cover classification. Photogram Eng Remote Sens. 1997;63:403–14.Search in Google Scholar
[53] Clark ML, Aide TM, Grau HR, Riner G. A scalable approach to mapping annual land cover at 250 m using MODIS time series data: A case study in the Dry Chaco ecoregion of South America. Remote Sens Environ. 2010;114:2816–32.10.1016/j.rse.2010.07.001Search in Google Scholar
[54] Tung F, LeDrew E. The determination of optimal threshold levels for change detection using various accuracy indexes. Photogram Eng Remote Sens. 1988;54:1449–54.Search in Google Scholar
[55] Janssen LL, Vanderwel FJ. Accuracy assessment of satellite derived land-cover data: a review. Photogram Eng Remote Sensing (United S). 1994;60:426–79.Search in Google Scholar
[56] Zhao Y, Gong P, Yu L, Hu L, Li X, Li C, et al. Towards a common validation sample set for global land-cover mapping. Int J Remote Sens. 2014;35:4795–814. 10.1080/01431161.2014.930202 Search in Google Scholar
[57] Fritz S, McCallum I, Schill C, Perger C, See L, Schepaschenko D, et al. An online platform for improving global land cover. Environ Model Softw. 2012;31:110–23.10.1016/j.envsoft.2011.11.015Search in Google Scholar
[58] Jafari M, Maghsoudi Y, Zoej MJV. A new method for land cover characterization and classification of polarimetric SAR data using polarimetric signatures. IEEE J Sel Top Appl Earth Observ Remote Sens. 2015;8:3595–607.10.1109/JSTARS.2014.2387374Search in Google Scholar
[59] Villa P, Stroppiana D, Fontanelli G, Azar R, Brivio PA. In-season mapping of crop type with optical and X-band SAR data: A classification tree approach using synoptic seasonal features. Remote Sens. 2015;7:12859–86.10.3390/rs71012859Search in Google Scholar
[60] Balzter H, Cole B, Thiel C, Schmullius C. Mapping CORINE land cover from Sentinel-1A SAR and SRTM digital elevation model data using random forests. Remote Sens. 2015;7:14876–98.10.3390/rs71114876Search in Google Scholar
[61] Yan WY, Shaker A, El-Ashmawy N. Urban land cover classification using airborne LiDAR data: A review. Remote Sens Environ. 2015;158:295–310.10.1016/j.rse.2014.11.001Search in Google Scholar
[62] Teluguntla P, Thenkabail PS, Oliphant A, Xiong J, Gumma MK, Congalton RG, et al. A 30-m landsat-derived cropland extent product of Australia and China using random forest machine learning algorithm on Google Earth Engine cloud computing platform. ISPRS J Photogram Remote Sens. 2018;144:325–40.10.1016/j.isprsjprs.2018.07.017Search in Google Scholar
[63] Gumma MK, Thenkabail PS, Deevi KC, Mohammed IA, Teluguntla P, Oliphant A, et al. Mapping cropland fallow areas in Myanmar to scale up sustainable intensification of pulse crops in the farming system. GI Sci Remote Sens. 2018;55:926–49.10.1080/15481603.2018.1482855Search in Google Scholar
[64] Sharma RC, Tateishi R, Hara K, Iizuka K. Production of the Japan 30-m land cover map of 2013–2015 using a Random Forests-based feature optimization approach. Remote Sens. 2016;8:429.10.3390/rs8050429Search in Google Scholar
[65] Friedl MA, Sulla-Menashe D, Tan B, Schneider A, Ramankutty N, Sibley A, et al. MODIS Collection 5 global land cover: Algorithm refinements and characterization of new datasets. Remote Sens Environ. 2010;114:168–82.10.1016/j.rse.2009.08.016Search in Google Scholar
[66] Defourny P, Kirches G, Brockmann C, Boettcher M, Peters M, Bontemps S, et al.: Product user guide version 2. 2018. https://www.esa-landcover-cci.org/?q=webfm_send/84 (accessed on 4 May 2019).Search in Google Scholar
[67] Li C, Gong P, Wang J, Zhu Z, Biging GS, Yuan C, et al. The first all-season sample set for mapping global land cover with Landsat-8 data. Sci Bull. 2017;62:508–15.10.1016/j.scib.2017.03.011Search in Google Scholar
© 2022 Junmei Kang et al., published by De Gruyter
This work is licensed under the Creative Commons Attribution 4.0 International License.
