Introduction

Arid zone grassland is one of the most vulnerable ecosystems on the planet, facing multiple challenges from global climate change and human activities. It plays an important role in food supply [1], soil conservation [2], water conservation and biodiversity conservation [3]. In addition, grassland resources have substantial cultural and economic value for local communities. Over the last few decades, overgrazing [4], agricultural expansion [5], urbanisation [6] and mining [7] have caused large changes in land cover (LC). An accurate LC classification map is important basic data for land managers and scientific researchers [8]. LC maps have been used extensively for research in land management, agricultural monitoring, ecosystem service assessment and climate change assessment [9]. Accurate, timely information on grassland area inventory is important for local government authorities, which provides a reliable basis for the rational development of grassland ecosystems and the macroscopic management of grassland resources.

Traditionally, grassland resource survey is archived by substantial field survey and measurement, which can accurately collect grassland spatial distribution and vegetation cover data. However, extensive fieldwork poses an immense challenge to human resources and is often labour intensive and time consuming [10]. With the rapid development of remote sensing technology and the free availability of satellite images, the scientific community has a unique opportunity to conduct mapping research in natural grasslands [11]. Currently, the available LC products have achieved remote sensing classifications on global [12, 13], national [14, 15] and regional [16, 17] scales at 30 m resolution. The producer accuracy of grassland ranges from 34.88% [12] to 88.01% [15] depending on the scale of study. The variation in producer accuracy for grassland is easily affected by the spatial heterogeneity of grassland [18]. Hou et al. [18] compared and analysed the consistency and accuracy of ten LC products, and concluded China urgently needs to develop a new grassland map. To validate the grassland accuracy of three existing land cover datasets (Chinese Academy of Sciences land use data (CNLUCC) [15] in 2018, Wuhan University annual land use data in China (WHCLCD) [14] in 2018, and GlobeLand30 [13] in 2020) in Ordos, 1,000 independent samples of visual interpretation were used to construct an error matrix, which revealed that the F1 scores for grassland ranged from 62.06% to 74.18%. However, it was observed that the F1 scores for grasslands were slightly low in the context of local scale study. Therefore, the utilization of high spatio-temporal resolution remote sensing satellite images for the refined extraction of grassland in arid and desertified regions is valuable and meaningful.

Optical remote sensing images are widely used in grassland mapping and monitoring [4]. Liu et al. [19] used Gaofen-1 satellite normalised difference vegetation index (NDVI) time-series data set with 8 m resolution to extract the spatial distribution information of the arid alfalfa artificial grassland in Linxi County, China, and the grassland patch accuracy reached 89.47%. However, the method has great limitations in natural grassland recognition and extraction due to the regular shape of the artificial grassland boundary and the fixed harvesting period. Herrero et al. [20] found that in discrete classification based on Landsat 8 OLI imagery with 30 m resolution, the support vector machine (SVM) algorithm and spectral features cannot be performed in southern African savanna land classification, and the classification accuracy was only 34.48%. The classification accuracy reached 79.31% after adding multitemporal NDVI and black decay temperature data to the classification features. The research of Herrero et al. [20] showed the increase in classification features’ dimensionality helps improve classification accuracy. The above studies were conducted for different types of grasslands, and the results showed remote sensing imagery with high spatial resolution can help improve mapping accuracy.

The quality of optical remote sensing images cannot be guaranteed because they are susceptible to the interference of clouds and rain [7, 21], which limits the remote sensing mapping of grassland. Compared with optical imagery, synthetic aperture radar (SAR) imagery has the advantage of all-day, all-weather acquisition and can be used as a supplementary data source to optical imagery. However, the signal of SAR image data is susceptible to interference from coherent speckle noise, which affects the identification of target features [22], and accurately extracting features from single-temporal-phase SAR data is difficult. Taravat et al. [23] explored the feasibility of the multitemporal SAR imagery of Sentinel-1 for the automatic detection of grass mowing status, with an overall accuracy of 85.71% for the validation set. The research result of Taravat et al. indicated the feasibility of using multitemporal SAR imagery for grassland growth state monitoring. Shafizadeh-Moghadam et al. [24]confirmed that the use of multi-temporal data improves the accuracy of land cover mapping. In the study of multi-temporal winter wheat mapping, Li et al. [25] utilised the NDVI curves of Modis images to segment the time window of synthetic images, and the results showed that the optimal selection of the time window helps to improve the accuracy of winter wheat mapping. The NDVI curve is useful for identifying the optimal time window for winter wheat mapping, but NDVI is susceptible to interference from soil background information in arid desert grasslands. To address this issue, four vegetation indices (NDVI, MSAVI, NDWI, and NDBI) were used to filter the time window and reduce interference from soil background information. Synthesising multi-temporal images using the optimal time window can help to address the issue of information loss that occurs when using a single synthesised image.

Optical remote sensing images and SAR images have their advantages, and their effective combination can improve the accuracy of grassland mapping, which has been successfully applied to grassland spatial distribution information extraction. Hong et al. [26] combined optical and SAR images to extract the spatial distribution of grassland and alfalfa in the southern prairie region of Saskatchewan, Canada, using unsupervised classification. In (SAR HV+MODIS) data combination mode, the grasslands’ user accuracy was highest at 89.7%, and that of alfalfa was 71.4%. In a single MODIS data model, the user accuracies of grasslands and alfalfa were 76.9% and 71.4%, respectively. Samrat et al. [27] extracted fragmented grassland patches in the tropics and found the overall accuracy was 96.32% for the combined-image classification, 91.93% for Sentinel-2 image and 63.23% for Sentinel-1 image. The above research results show the integrated images have different degrees of improvement in the accuracy of grassland mapping compared with single optical remote sensing images and SAR images. By contrast, de Oliveira Santos et al. [28] used seasonal multisensor image time series for the classification of crops, pastures and afforestation in subtropical agricultural regions, and found the addition of Sentinel-1 images does not improve the classification accuracy. Based on the literature review, the integration of optical images and SAR images is proposed as a solution to enhance the extraction accuracy of natural grasslands in arid areas, leveraging their respective advantages.

Currently, the choice of ML algorithms directly influences the grassland mapping accuracy. In a study comparing random forest (RF), SVM, decision trees (DTs), and artificial neural networks (ANNs) classifiers for Landsat-based LULC mapping, Shih et al. [29] discovered that RF had the highest average overall accuracy (OA) of 71.50%. Furthermore, RF required less computing time compared to the other classifiers. In Heydari’s study on mapping large areas of land cover using deep neural networks and Landsat time-series observations [30], it was observed that deep neural network algorithms do not enhance the accuracy of mapping solely through spectral feature classification when compared to traditional SVM algorithms. Furthermore, the study found that deep neural networks necessitate more intricate parameter settings. RF is known for robustness to noise and their ability to handle overfitting, thus having strong generalizability and transferability [31]. However, RF algorithm is easily affected by the unbalanced ratio between training sample classes. RF classifiers are constructed by minimising the overall classification error, and unbalanced training samples can lead to poor accuracy for a few classes [32]. The class imbalance problem occurs when one or some classes have fewer samples than others, and learning from unbalanced training data is a common problem in machine learning [8]. In Ordos, the proportion of grassland area reaches more than 50%, and the imbalance between grassland and other land cover types is unavoidable. To alleviate the problem of unbalanced training samples, scholars have explored techniques, such as undersampling of majority classes [33], oversampling of minority classes [34], and a combination of oversampling and undersampling training classes [35]. An incorrect sampling strategy affects the performance of the model, which increases training complexity and leads to overfitting problems [36]. Thus, the sampling strategy of unbalanced training samples must be investigated to improve the accuracy of LC types with small area proportions. To address unbalanced sampling during model training, Zhang et al. [37] proposed a random sampling strategy with the hexagonal tile model to improve LC mapping accuracy in Madagascar with highly heterogeneous landscapes. The producer accuracy of LC based on the traditional global classification model of the classifier was 84.8%, and the producer accuracy of LC based on the tile classification model was 88.2%. Inspired by the study of Zhang et al. [37], the rectangular tile classification model is proposed based on Sentinel integrated images in the study of grassland mapping in arid zones in anticipation of improving the accuracy of grassland mapping.

The Ordos Plateau is in the southwest of the Inner Mongolia Autonomous Region, which is a typical agro-pastoral zone and sensitive to climate change and human activities. From southeast to northwest, the grasslands of the Ordos Plateau are divided into typical grassland subzone, desert grassland subzone and grassland desertification subzone [38]. Its landscape features are highly heterogeneous and fragmented [39], which brings some difficulties to accurate grassland mapping in this region. To enhance the accuracy of Ordos grassland mapping and address the problem of unbalanced training samples, the rectangular tile classification model based on Sentinel integrated images is proposed in this paper. The research objectives of this paper are as follows: (1) validate the effectiveness of the rectangular tile classification model in improving the accuracy of grassland mapping in arid regions with highly heterogeneous landscapes, (2) determine whether the overall accuracy of grassland mapping in arid regions can be improved by combining Sentinel-1A and Sentinel-2 images and (3) determine the potential of vegetation index features and texture features in improving mapping accuracy.

Study area and data collection

Overview of the study area

The Ordos Plateau is in the southwest of Inner Mongolia, China. The regional latitude and longitude range are 106°42′ E–111°31′ E and 37°41′ N–40°51′ N, respectively (Fig 1). The study area is about 360 km long from north to south and 430 km wide from east to west, with a total area of 86,859 km² and an average altitude of 1,000–1,500 m. By the end of 2021, the resident population of Ordos was 2.168 million people, the regional gross domestic product (GDP) was 68.61 billion USD and the GDP per capita was 31,639 USD [40]. The natural geographical environment of the Ordos Plateau is unique. The topography is high in the west and low in the east. The region has a temperate continental arid climate. The average precipitation in the east is 300–400 mm, and that in the west is 190–300 mm. The annual potential evaporation is as high as 2,000–3,000 mm. The annual average temperature is 5.3°C–8.7°C, the average wind speed is 2.7–3.7 m/s and the annual gale days are about 40 d. Kubuqi Sand in the north and Maowusu Sand in the southcentral are the two main geomorphic units in the study area, accounting for 48% of the total area, with sparse vegetation and a harsh natural environment. The alluvial plain along the Yellow River in the north, which accounts for 4% of the total land area, is fertile and rich in agricultural and livestock products. The eastern part is a hilly, gully area with serious soil erosion, and the ecological environment is fragile, accounting for 19% of the total land area. The western part is an undulating plateau area, accounting for 29% of the total land area. It is a typical desert grassland with minimal rainfall [41].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study area.

https://doi.org/10.1371/journal.pone.0301444.g001

Data and preprocessing

Sentinel data and preprocessing.

The sentinel data used in this paper include Sentinel-1A and Sentinel-2, which are from the European Space Agency. The Sentinel-1A [42] active microwave remote sensing satellite carries a sensor of SAR in the C-band, which has a revisit cycle of 12 d. Ground range detected (GRD) products with a resolution of 10 m and polarisation of VV (vertical vertical) and VH (vertical horizontal) in Interferometric Wide Swath (IW) mode were selected as data sources. On the GEE platform, GRD products have completed preprocessing, such as radiation correction, thermal noise removal and terrain correction. Data loading can be performed by using only the relevant functions, which greatly reduces the difficulty of data processing. Sentinel-2 [43] consists of two high-resolution satellites, Sentinel-2A and Sentinel-2B, with a revisit period of 10 d for a single satellite. The two satellites complement each other, and the revisit cycle is reduced to 5 d. The two satellites are equipped with multispectral imagery (MSI) sensors with 13 bands, including visible, near-infrared (NIR) and short-wave infrared, with spatial resolutions of 10, 20 and 60 m. The reflectance product (MSI L1C image) with orthorectification and geometric correction was selected as the data source. The cloud amount of this image was less than 15%, and the QA60 band was used for cloud removal. A total of 934 Sentinel-2 MSI L1C images from May 9, 2018, to November 2, 2018, and 212 Sentinel-1A GRD images from May 1, 2018, to October 31, 2018, were used in this paper (Table 1). During data preprocessing, the 10 m resolution VV and VH bands of Sentinel-1A, and the 10 m resolution blue, green, red and NIR bands of the Sentinel-2 were resampled to 20 m.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Satellite image data.

https://doi.org/10.1371/journal.pone.0301444.t001

MODIS data.

MODIS images acquired by Terra satellite, which had a spatial resolution of 500 m and a temporal resolution of 8 d, were used in this paper [44]. The MODIS time-series image data were used to calculate the vegetation indices, and the optimal synthetic time window for Sentinel-2 images was determined by time-series vegetation index curves. The S–G filtering model [45] was used to smooth the anomalous noise of the time-series vegetation indices to obtain vegetation index curves that are more consistent with the grassland phenology. This method eliminates the influence of cloud cover and atmospheric factors on the time-series vegetation index data and facilitates the determination of more accurate image synthesis time windows. Thirty-eight MODIS images from March 6, 2018 to December 27, 2018, were used in this paper (Table 1).

Sample data.

Existing Chinese LC products were used for analysis to determine the major LC distribution in the study area, and sample data were randomly generated based on the results. LC types were classified into six major categories: grassland, cultivated land, woodland, waterbody, building and unused land. The 2018 Chinese Academy of Sciences land use data (CNLUCC) [15] and Wuhan University annual land use data in China (WHCLCD) [14] were reclassified based on LC types, and their spatially identical parts were manually clicked to generate 7,938 sample points, ensuring the uniformity of the spatial distribution of sample points. The sample types and numbers are shown in Table 2. The spatial distribution of all sample points is shown in Fig 2. The sample points were corrected by visual interpretation based on the GEE platform. The sample data were imported into the GEE platform. Approximately 70% and 30% of the sample data were used as training set for model construction and validation set for classification accuracy verification, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The spatial distribution of all sample points.

https://doi.org/10.1371/journal.pone.0301444.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Sample points of the study.

https://doi.org/10.1371/journal.pone.0301444.t002

Methodology

In this paper, the classification task was employed on the GEE platform (https://earthengine.google.com/). The basic flow of the research is shown in Fig 3. The monthly Sentinel-1A median value images were synthesised, and the optimal synthesis time window of Sentinel-2 images was determined by MODIS NDVI, modified soil adjusted vegetation index (MSAVI), normalised difference water index (NDWI) and normalised difference built-up index (NDBI) of mean value time-series curves for grassland (Methodology: Image Composition Solution). The polarisation features, spectral features, vegetation index features and texture features were constructed as classification feature sets (Methodology: Selection of Feature Variables). Seven experimental groups (Methodology: Experimental Scheme) were designed, including the rectangular tile classification model and the traditional global classification model. The spatial distribution map of Ordos grassland was drawn based on the rectangular tile classification model (Methodology: Rectangular Tile Classification Model) and the RF algorithm (Methodology: RF Algorithm), and its classification accuracy was evaluated (Methodology: Accuracy Evaluation).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Flow chart of the data process.

Note: Abbreviation: CNLUCC, Chinese Academy of Sciences land use data; WHUCLCD, Wuhan University annual land use data in China. The rectangular tile model represents the study area divided into rectangles of equal size, and the traditional global model represents the study area not divided into tiles.

https://doi.org/10.1371/journal.pone.0301444.g003

Image composition solution

Different image synthesis schemes were proposed for SAR images and optical images, considering the characteristics of different types of sensors. The key time period of grass growth (May to October 2018) was selected to synthesise monthly Sentinel-1A median value images for six periods.

Ensuring the acquisition of valid, high-quality Sentinel-2 images from May to October 2018 was difficult due to the influence of clouds, rain and sunlight. MODIS images were used to calculate four vegetation indices, NDVI, NDWI, NDBI and MSAVI, for determining the optimal time window for synthetic images. This process was performed to reduce data redundancy and enhance remote sensing information whilst ensuring the authenticity and reliability of the data. Four vegetation index mean value curves of NDVI, NDWI, NDBI and MSAVI were plotted to describe the characteristics of grassland changes over time (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. MODIS average NDVI, MSAVI, NDWI, NDBI time series curves for grassland.

Note: Abbreviation: NDVI, Normalized Difference Vegetation Index; MSAVI, Modified Soil Adjusted Vegetation Index; NDWI, Normalized Difference Water Index; NDBI, Normalized Difference Built-up Index.

https://doi.org/10.1371/journal.pone.0301444.g004

The maximum value images with a time window from June 10 to October 24 were synthesised by NDVI and MSAVI. The minimum value image with a time window from June 18 to November 2 was synthesised by NDWI. The minimum value image with a time window from June 10 to October 24 was synthesised by NDBI. If the Sentinel-2 images were synthesised with the time windows of the vegetation indices, then data redundancy would occur, and the data processing speed would be reduced. Therefore, two separate phases of optical median value images were synthesised based on the growth and senescence stages of the grass. The time windows were the growth period from May 9 to August 12 and the senescence period from August 13 to November 2.

Selection of feature variables

Five features, including the original spectral bands of optical images, vegetation index features and texture features, and the polarisation bands and texture features of SAR images, were selected in this paper (Table 3). The visible band, NIR band and short-wave infrared band of Sentinel-2 images are widely used in vegetation remote sensing, which are of great value for vegetation identification. NDVI plays a crucial role in the classification of vegetation classes [46] and can be used to categorise bare sand and open waterbody very well [47]. MSAVI is very important to increasing the vegetation signal and reduce soil-induced changes simultaneously, thus minimising the effect of soil on vegetation spectra [48]. NDWI provides valuable information on classification characteristics by considering the variability of vegetation water content [49]. NDBI is effective in distinguishing information about vegetation, water bodies and buildings [50].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Feature variables of the satellite data.

https://doi.org/10.1371/journal.pone.0301444.t003

Texture can present the surface or structural properties of an image and is an important feature variable to improve the accuracy of LC classification. However, many texture features cause data redundancy. This paper filtered the texture features whilst retaining the maximum amount of information. Angular second moment (ASM), contrast (CON), correlation (COR) and entropy (ENT) [22] were selected as characteristic variables. ASM represents the coarseness of the texture and the uniformity of grey distribution, CON represents the depth of the texture grooves and the clarity of the image, COR represents the consistency of the texture in the local area and ENT represents the complexity of the texture.

Experimental scheme

Seven experimental groups were designed, including 14 experimental schemes (Table 4). Single optical imagery and single SAR imagery were designed to complete the grassland distribution information extraction (EXP1-EXP5). This process was performed to investigate the feasibility and the effectiveness of various types of sensor images for grassland mapping. The overall accuracy of the mapping of Sentinel-integrated images (EXP6-EXP7) was compared with the experimental results of the classified mapping of Sentinel-1A SAR images and Sentinel-2 optical images to explore the potential of the integrated images of SAR and optical images in improving the accuracy of grassland mapping.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Experimental design.

https://doi.org/10.1371/journal.pone.0301444.t004

LC in the highly heterogeneous areas of the landscape exhibited complex, fragmented characteristics with changes in geospatial location, especially in the study area where geographic longitudes and latitudes span large distances. The study area was divided into 15 tiles (Fig 5), and LCs in each tile possessed similar spectral characteristics to address the effects of geospatial location differences on spectral features. This method obeys the first law of geography and can effectively reduce the identification error caused by different spectral characteristics and phenological characteristics of the same LC type in the study area.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Schematic of the random forest classifier operation based on the rectangular tile classification model.

Note: First Step: 0–15 represents the coding order of the tiles. Second Step: The blue tile represents the central tile and the white tiles represent the adjacent tiles.

https://doi.org/10.1371/journal.pone.0301444.g005

Rectangular tile classification model

In this paper, the study area was divided into 15 equal-area rectangular tiles, the images within each tile were independently classified based on the RF algorithm, and all tile classification results were integrated into a final classification result map (Fig 6). During the model training, the tile order was cycled sequentially. The rectangular tile classification model utilised the training samples of the central tile and the surrounding adjacent tiles to participate in the RF model training jointly. Compared with the rectangular tile classification model, the traditional global classification model utilised all the training samples in the study area as the construction of the RF algorithm model at the same time. The tile classification model reduced the error of the variation of spectral features amongst different regions [37] and ensured the random uniform distribution of different LCs in Ordos. Sample points of approximately six major LC types within each rectangular tile were randomly selected as training samples. However, each tile had a different number of samples within it to balance the proportion of small area LC types, such as cultivated land, forest land, waterbody and building. All samples were selected uniformly across the study area to avoid spatial correlation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Comparisons among EXP2, EXP5, and EXP7 by F1 score.

Note: EX2: S1 Polarization feature + S1 Texture feature; EXP5: S2 Spectral feature + S2 Vegetation index + S2 Texture feature; EXP7: S1 Polarization feature + S1 Texture feature + S2 Spectral feature + S2 Vegetation index + S2 Texture feature. The F1 score is a weighted average of the mapping accuracy and user accuracy.

https://doi.org/10.1371/journal.pone.0301444.g006

Results

Accuracy of experimental schemes

The grassland distribution information was extracted by using different experimental schemes, and the OA, kappa coefficient, PA, UA and F1 score of different schemes were calculated. The results in Table 5 show OA improved with the increase in the number of feature variables when using Sentinel-1A and Sentinel-2 images alone for grassland mapping. Based on the rectangular tile classification model, the F1 Scores showed a slight increase as the dimensionality of the classification features increased, which can be observed when comparing EXP3-2, EXP4-2, EXP5-2, EXP6-2, and EXP7-2. In particular, the F1 score in EXP7-2 on the rectangular tile classification model reached the maximum of 89.24%. When using Sentinel-integrated images to extract grassland, the OA was improved in varying degrees compared with that of a single sensor image. For each experimental group, all five grassland mapping accuracy metrics of the rectangular tile classification model were improved to a greater extent compared with the traditional global classification model. Comparisons of the mapping results of EXP2, EXP5 and EXP7 based on the traditional global classification model and the rectangular tile classification model were evaluated by F1 scores (Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Accuracy of experimental schemes.

https://doi.org/10.1371/journal.pone.0301444.t005

The detailed comparisons of three LULC maps between EXP2, EXP5 and EXP7 are shown in Fig 7. Two complex typical small regions were segmented for comparative analysis based on the traditional global classification model and the rectangular tile classification model. Region 1 (110°42′33″ E, 39°36′35″ N) is in the eastern gully area of the study area and is a typical grassland zone. Region 2 (109°43′08″ E, 39°33′05″ N) is in the urban area of Ejin Horo Banner. In region 1, the classification results of EXP2 and EXP5 had more building speckle noise based on the rectangular tile and traditional global classification models, but the classification results of EXP7 had considerably less building speckle noise. In EXP7, the classification results based on the rectangular tile classification model were clearer and neater, and the whole area was more consistent with the actual LC. In Region 2, the classification results of EXP2 clearly showed the misclassification of buildings, cultivated lands and grasslands. EXP5 had a part of cultivated land and grassland patches misclassified as unused land based on the rectangular tile and traditional global classification models, and the classification results of EXP7 were improved to some extent. The classification results of EXP7 were based on the rectangular tile classification model, and the classification results within the whole area were more consistent with the actual LC.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Detailed comparisons among classification result from EXP2, EXP5, and EXP7 by spatial patterns in two regions.

https://doi.org/10.1371/journal.pone.0301444.g007

Grassland spatial distribution

From EXP7-2 based on Sentinel-integrated images, the spatial distribution map of Ordos grassland with a spatial resolution of 20 m was drawn by using the rectangular tile classification model to extract the grassland distribution information of the study area. Comparisons between the classification result from EXP7-2, CNLUCC-2018 (Kuang et al., 2022) and WHCLCD-2018 (Yang and Huang, 2021) by spatial patterns indicate most of the grasslands were correctly classified, grasslands in the region were mainly distributed in the central and south-western parts of the study area, and the vegetation was relatively sparse in the Kubuqi Sand in the north and the Maowusu Sand in the southeast.

The error matrix from EXP7-2 was constructed by using the classification results of grassland, woodland, cultivated land, waterbody, building and unused land in the study area. The OA, kappa coefficient, PA, UA and F1 score were calculated. The results are shown in Table 6. The PA, UA and F1 score of the grassland were 93.11%, 85.68% and 89.24% respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Confusion matrix for EXP7-2 of the grassland distribution map in Ordos, China.

https://doi.org/10.1371/journal.pone.0301444.t006

Comparisons between the Derived LULC map, CNLUCC-2018 and WHCLCD-2018 of Ordos

The obtained 20 m Sentinel-2 derived LULC map from EXP7-2 in this paper and the 30 m Landsat-derived LULC maps (CNLUCC-2018 and WHCLCD-2018) were compared in terms of F1 score (Fig 8). Prior to the comparative analysis, CNLUCC-2018 and WHCLCD-2018 were reclassified into a unified LC criteria system. The F1 score for grassland was lower than CNLUCC-2018 and higher than WHCLCD-2018.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Comparisons among classification result from EXP7-2, CNLUCC-2018, and WHCLCD-2018 by F1 score.

https://doi.org/10.1371/journal.pone.0301444.g008

The detailed comparisons of three LULC maps are shown in Fig 9. Four typical small regions were segmented from the study area for detailed comparative analysis. Region 3 (110°56′11″ E, 39°48′53″ N) is in the eastern gully area of the study area and is a typical grassland zone. Region 4 (109°18′51″ E, 40°21′22″ N) is in the northern part of the study area in the Kubuqi Sand, which is a desert grassland zone. Region 5 (107°58′10″ E, 38°08′04″ N) is in the south-eastern part of the study area in the Maowusu Sand, which is the desert grassland zone. Region 6 (107°24′10″ E, 39°39′13″ N) is in the undulating plateau zone in the western part of the study area, which is the grassland desertification zone. In region 3, our results were more consistent with the actual LC, and the red line feature roads were correctly extracted. The road information in CNLUCC-2018 and WHCLCD-2018 was not shown. The area of unused land and building in CNLUCC-2018 was large. The woodland area in WHCLCD-2018 was small. In region 4, our results were more consistent with WHCLCD-2018, where fine-grazed grassland patches and cultivated land patches in the sand were correctly extracted. A larger red building patch was clearly misclassified in CNLUCC-2018, and a larger number of cultivated land patches were omitted. In region 5, our results and WHCLCD-2018 were more consistent with the actual LC. The area of unused land in CNLUCC-2018 was remarkably larger, and all cultivated patches were omitted. In region 6, our results were more consistent with the actual LC. However, one red building patch was missed. The red building patch in CNLUCC-2018 was extracted correctly, but many fine unused land patches were missed. The red building patch in WHCLCD-2018 was also missed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Detailed comparisons among classification result from EXP7-2, CNLUCC-2018, and WHCLCD-2018 by spatial patterns in four regions.

https://doi.org/10.1371/journal.pone.0301444.g009

Discussion

Comparing the grassland mapping accuracy of EXP3-2, EXP4-2, and EXP5-2 by F1 score, our results showed that the integration of vegetation index features and texture features improved the accuracy of grassland mapping. In EXP4-2, the inclusion of vegetation indices bands resulted in a 0.75% improvement in grassland mapping accuracy compared to EXP3-2. Similarly, in EXP5-2, the addition of texture features led to a further improvement of 0.84% in grassland mapping accuracy compared to EXP4-2. Two main reasons explain the difficulty of extracting information from grasslands in Ordos arid zone: Firstly, shrubs and herbs are the main vegetation types, vegetation cover is low, and vegetation is scattered in the arid desert grassland subzone and grassland sandy subzone [55]. Secondly, bare sand within unused land has a substantial effect on the spectral behaviour of the canopy of shrub and herbaceous vegetation with sparse vegetation cover [56]. To reduce the adverse effects of low vegetation cover and bare sand, Ge et al. [47] introduced NDVI, MSAVI, NDWI and NDBI indices to improve the accuracy of LC mapping in arid desert oases effectively. Shafizadeh-Moghadam et al. [24] found that incorporating texture features could be beneficial in enhancing overall accuracy in study of LC mapping of six Middle Eastern countries in the Tigris-Euphrates River Basin.

Integration of Sentinel-1A SAR imagery to Sentinel-2 images can improve the accuracy of grassland mapping. Comparing the grassland mapping accuracy of EXP5-2 and EXP7-2, the mapping accuracy of Sentinel-integrated images was improved by 1.66% compared with Sentinel-2 images in terms of F1 score. Previous studies [27, 57] showed the use of integrated imagery to map the spatial distribution of grasslands is more advantageous. Their research results showed integrated images help improve the accuracy of LC classification in grassland areas. When optical images are affected by clouds and rain resulting in less optical image coverage in the study area, it can greatly limit the remote sensing land classification mapping in the study area. In this case, SAR images can be used as supplementary data for optical images because the microwave of the SAR sensor can penetrate the canopy of vegetation to obtain information on vegetation coverage, vegetation moisture content and soil background. In the Ordos Plateau, cloudy weather is concentrated from February 9 to September 15 every year. Therefore, LC classification using Sentinel-integrated images in this paper is feasible and valuable.

In our paper, the grassland mapping accuracy of all experimental groups based on the rectangular tile classification model was higher than that of the traditional global classification model. For large-scale LC mapping, a certain type of LC can easily be mistaken for other types of LC due to the influence of spectral features, for example, woodland and grassland [37]. Domestic and international scholars have recently conducted some research to improve the accuracy of remote sensing mapping on large scale. Bartalev et al. [58] proposed a new locally adaptive classification method (locally adaptive global mapping algorithm) for large-scale LC mapping using remote sensing data, which was the grid-based supervised image classification using classes’ features estimated locally in classified pixels’ surrounding from spatially distributed reference data. They found the traditional global classification model accuracy was 81.8%, and the local model accuracy was 96.8%. Although the classification algorithm improved the classification accuracy remarkably, the limited data processing capacity of the local platform affected the mapping efficiency.

The emergence of GEE provides a feasible means to overcome this challenge. GEE greatly facilitates researchers worldwide to conduct research on large-scale LC classification and dynamic monitoring of land use change. Zhang et al. [37] used Sentinel-2 time series, tile-based image classification and GEE to generate high-resolution LC in Madagascar automatically. They found the average PA of the traditional global classification method was 84.8%, and the PA of the tile classification method was only 88.2%, which indicated the tile classification method can remarkably improve the accuracy of classification mapping and reduce misclassification amongst LC types. Xuan et al. [59] mapped crop type in Northeast China during 2013–2021 using automatic sampling and tile-based image classification, but they did not conduct a comparative analysis of the classification accuracy of the tile classification model and the traditional global classification model. The tile classification model based on RF classifier considered the similarity and specificity of vegetation structure within climatic subzones. Inspired by the above studies, our remote sensing mapping study of Ordos grassland was based on the rectangular tile classification model and integrated image on the GEE platform. In our paper, the F1 score of grassland in EXP7-2 on the rectangular tile classification model reached the maximum value of 89.24%, which was improved by 13.02% compared with EXP7-1 on the traditional global classification model. The adoption of the rectangular tile classification model in large-scale LC classification mapping was conducive to remarkably improving the mapping accuracy. This conclusion was consistent with the findings of Zhang et al. [37] in LC mapping in Madagascar, which provided a new idea for subsequent LC mapping of highly heterogeneous landscapes on a large scale.

Limitations and recommendation

Although the accuracy of grassland mapping based on the rectangular tile classification model was remarkably improved than the traditional global classification model, the paper had limitations. Firstly, the optimal scale regarding the rectangular tile classification model was not explored in depth. Only rectangular tiles were utilised, and other shapes of tiles were ignored. Secondly, the spectral differences between images were not explored during image synthesis, which may increase the error of the experimental results because the classifier treats excessive differences as noise. Thirdly, Sentinel-2 images using time window synthesis can reduce data redundancy, but the information of Sentinel-2 time-series images cannot be fully exploited. Sentinel-2 monthly synthetic images can be considered to enhance the utilisation of image information. Finally, only one algorithm of RF was utilised. In future study, we will focus on enhancing the construction of sample datasets and vegetation index feature datasets. Additionally, we will strive to improve the algorithms by incorporating soft classification algorithms and deep learning algorithms into the study of grassland mapping in arid zones.

Conclusion

In this paper, the rectangular tile classification model based on RF algorithm and Sentinel-integrated imagery on GEE platform was utilised to map the spatial distribution of Ordos grassland in 2018. Comparing the grassland mapping accuracy of EXP3-2, EXP4-2, and EXP5-2 by F1 score, our results showed that the integration of vegetation index features and texture features improved the accuracy of grassland mapping. Integration of Sentinel-1A SAR imagery with Sentinel-2 images can improve the accuracy of grassland mapping. In all seven experimental groups, the rectangular tile classification model was found to improve overall accuracy (OA) by 1.20% to 13.99% compared to the traditional global classification model. The adoption of the rectangular tile classification model in large-scale grassland mapping was conducive to remarkably improving the mapping accuracy in arid zone. This paper and its results provide some new viewpoints and guidance for remote sensing mapping in highly heterogeneous landscapes areas on a large scale.

Supporting information

Acknowledgments

The authors are sincerely grateful to the Resource and Environment Science and Data Center, Chinese Academy of Sciences and Professor Xin Huang’s group for selflessly sharing their LULC products. We offer our special thanks to Mr. Zhang of GEE Studyroom for his selfless sharing of the tile classification model code and the Google team for building the powerful Google Earth Engine (GEE) platform.

For more details about supporting information, readers are referred to the online material (https://figshare.com/s/59a7839937a5b10dde1d).