Mapping of Kobresia pygmaea Community Based on Umanned Aerial Vehicle Technology and Gaofen Remote Sensing Data in Alpine Meadow Grassland: A Case Study in Eastern of Qinghai–Tibetan Plateau

: The Kobresia pygmaea (KP) community is a key succession stage of alpine meadow degradation on the Qinghai–Tibet Plateau (QTP). However, most of the grassland classiﬁcation and mapping studies have been performed at the grassland type level. The spatial distribution and impact factors of KP on the QTP are still unclear. In this study, ﬁeld measurements of the grassland vegetation community in the eastern part of the QTP (Counties of Zeku, Henan and Maqu) from 2015 to 2019 were acquired using unmanned aerial vehicle (UAV) technology. The machine learning algorithms for grassland vegetation community classiﬁcation were constructed by combining Gaofen satellite images and topographic indices. Then, the spatial distribution of KP community was mapped. The results showed that: (1) For all ﬁeld observed sites, the alpine meadow vegetation communities demonstrated a considerable spatial heterogeneity. The traditional classiﬁcation methods can hardly distinguish those communities due to the high similarity of their spectral characteristics. (2) The random forest method based on the combination of satellite vegetation indices, texture feature and topographic indices exhibited the best performance in three counties, with overall accuracy and Kappa coefﬁcient ranged from 74.06% to 83.92% and 0.65 to 0.80, respectively. (3) As a whole, the area of KP community reached 1434.07 km 2 , and accounted for 7.20% of the study area. We concluded that the combination of satellite remote sensing, UAV surveying and machine learning can be used for KP classiﬁcation and mapping at community level.


Introduction
Alpine meadow is the major vegetation type on the Qinghai-Tibet Plateau (QTP), China. It is important for animal husbandry, water conservation and biodiversity conservation [1,2]. Since the 1980s, due to the dual effects of climate change and human activities, alpine meadow grassland has experienced different extents of degradation, especially in In this study, we aimed to map the KP community over the eastern of QTP by using the combination of UAV aerial photographing, GF WFV images and machine learning algorithms. We hope this study can be helpful for guiding further mapping of the KP community over the whole QTP and provide scientific basics for restoration and management activities.

Study Area
The study area is located at the eastern of the source region of the Yellow River, including Zeku County and Henan County of Qinghai province, and Maqu County of Gansu province (Figure 1). It is one of the most important animal husbandry basis on the QTP and also an important water source conservation area in China. The study area is located at 33°03′~35°33'N, 100°33'~102°33'E, with elevation ranging from 2871 to 4850 m ( Figure 1c). The mean annual precipitation ranges from 400~600 mm, mean annual temperature is between −2.4~2.1 °C It belongs to the continental plateau temperate monsoon climate. Alpine meadow is one of the main alpine grassland types, accounting for 79.67% of the whole study area. Other than alpine meadow, mountain meadow, swamp meadow and alpine steppe account for 13.22%, 1.78%, and 1.69%, respectively ( Figure 1b). The growth period of grassland plants is relatively short, only about 150 days, mainly from May to September. The grasslands are mainly used for yak and sheep grazing.

Field Observation and Preprocess of Aerial Photographs
We carried out the field monitoring for vegetation communities of alpine meadow based on aerial photographs by Phantom 3 professional and Mavic 2 zoom Quad-Rotor intelligent UAVs (manufactured by DJI Innovation Industries; http://www.dji.com (accessed on 2018). According to grassland growth status and spatial representativeness, an area in the range of 250 × 250 m was selected as an observation site, and four flight routes were designed in each site, including one GRID flight way (200 × 200 m) and three BELT flight ways (40 × 40 m) (Figure 2a). The flight way of UAVs was designed by

Field Observation and Preprocess of Aerial Photographs
We carried out the field monitoring for vegetation communities of alpine meadow based on aerial photographs by Phantom 3 professional and Mavic 2 zoom Quad-Rotor intelligent UAVs (manufactured by DJI Innovation Industries; http://www.dji.com (accessed on 1 June 2018). According to grassland growth status and spatial representativeness, an area in the range of 250 × 250 m was selected as an observation site, and four flight routes were designed in each site, including one GRID flight way (200 × 200 m) and three BELT flight ways (40 × 40 m) (Figure 2a). The flight way of UAVs was designed by FragMAP [22], Phantom 3 professional was used to perform the GRID flight way at a height of 20 m (red dot in Figure 2a,b), Mavic 2 zoom was used to perform the BELT flight way at a height of 2 m (green dot in Figure 2a,c). The positional accuracy of two UAVs was ±1.5 m horizontally and ±0.5 m vertically. A photograph was then taken vertically downward at each way point automatically, the photograph resolutions of GRID and BELT were 1 and 0.09 cm, and the ground coverages were 26 × 35 m and 2.57 × 3.43 m, respectively. flight way at a height of 2 m (green dot in Figure 2a,c). The positional accuracy of two UAVs was ±1.5 m horizontally and ±0.5 m vertically. A photograph was then taken vertically downward at each way point automatically, the photograph resolutions of GRID and BELT were 1 and 0.09 cm, and the ground coverages were 26 × 35 m and 2.57 × 3.43 m, respectively.
To better identify the vegetation species, about 9~15 aerial photographs were collected randomly by operating Mavic 2 zoom manually at a height of 0.5 m in each sample site. The number of photographs was determined by the uniformity of community growth status. These aerial photographs could clearly identify plant species, which was corresponding to the traditional ground observation quadrat (Supplementary Figure S1). According to the dominant species of grass vegetation, grassland coverage, texture features and plateau pika (Ochotona curzoniae, hereafter pika) activities, the aerial photographs were divided into six types, including four alpine meadow vegetation communities of Poaceae, KH, KP and BS ( Figure 3 and Table 1), two land covers of shrub meadow (SM) and marsh meadow (MM). Additionally, the forest and others (bare land, construction use and waters) were acquired based on the Google Earth images and GF WFV images. Field observation was carried out at the peak time of grassland growth, and 751 sample sites were observed from 2015 to 2019 in total (Figure 1c). About 30 sample sites were acquired for forest and others. Grassland had a unique morphology and textural characteristics, with closed and monospecific builds (2~3 cm in height), polygonal crack patterns and a felty root mat, pika and poisonous weeds To better identify the vegetation species, about 9~15 aerial photographs were collected randomly by operating Mavic 2 zoom manually at a height of 0.5 m in each sample site. The number of photographs was determined by the uniformity of community growth status. These aerial photographs could clearly identify plant species, which was corresponding to the traditional ground observation quadrat (Supplementary Figure S1).
According to the dominant species of grass vegetation, grassland coverage, texture features and plateau pika (Ochotona curzoniae, hereafter pika) activities, the aerial photographs were divided into six types, including four alpine meadow vegetation communities of Poaceae, KH, KP and BS ( Figure 3 and Table 1), two land covers of shrub meadow (SM) and marsh meadow (MM). Additionally, the forest and others (bare land, construction use and waters) were acquired based on the Google Earth images and GF WFV images. Field observation was carried out at the peak time of grassland growth, and 751 sample sites were observed from 2015 to 2019 in total ( Figure 1c). About 30 sample sites were acquired for forest and others.

Region of Interest Construction
According to the GPS information recorded in FragMAP and stored in aerial photographs property files, the names of photographs were renamed by the number of 1 to 16 by the DJI Locator software [22] in each site. Then the region of interest was built based on photograph location information in the same observation site in ArcGIS and ENVI software (Figure 3d,h,l,p). Additionally, about 30 samples (region of interest, ROI) for forest and others (the water, bare land, and construction land) were selected in ENVI software, according to the GF WFV images and Google Earth images.

Region of Interest Construction
According to the GPS information recorded in FragMAP and stored in aeria photographs property files, the names of photographs were renamed by the number of to 16 by the DJI Locator software [22] in each site. Then the region of interest was buil based on photograph location information in the same observation site in ArcGIS and ENVI software (Figure 3d,h,l,p). Additionally, about 30 samples (region of interest, ROI for forest and others (the water, bare land, and construction land) were selected in ENV software, according to the GF WFV images and Google Earth images.

Acquisition and Preprocessing of Remote Sensing Data
The remote sensing data, including GF1 and GF6 WFV imager images, wer downloaded from the China Centre for Resources Satellite Data and Application (http://www.cresda.com/EN/ (accessed on 2019 and 2018)). The WFV imager was carried by GF1 and GF2 satellites, with four multi-spectral bands (800 km of swath width) and eight multi-spectral bands (850 km of swath width), respectively. The resolution of WFV image was 16 m, and the revisit period for each satellite was 4 days. Together, the revisi period could be reached up to 2 days (Table 2). Three scenes of WFV images with no cloud cover in Zeku, Henan and Maqu, during the peak of grassland growth of 2019 and 2020 were downloaded ( Table 3). The GF WFV data were preprocessed using ENVI 5.

Acquisition and Preprocessing of Remote Sensing Data
The remote sensing data, including GF1 and GF6 WFV imager images, were downloaded from the China Centre for Resources Satellite Data and Application (http://www. cresda.com/EN/ (accessed on 20 September 2019)). The WFV imager was carried by GF1 and GF2 satellites, with four multi-spectral bands (800 km of swath width) and eight multi-spectral bands (850 km of swath width), respectively. The resolution of WFV image was 16 m, and the revisit period for each satellite was 4 days. Together, the revisit period could be reached up to 2 days (Table 2). Three scenes of WFV images with no cloud cover in Zeku, Henan and Maqu, during the peak of grassland growth of 2019 and 2020 were downloaded ( Table 3). The GF WFV data were preprocessed using ENVI 5.3 software, and the Radiometric Calibration module, FLAASH Atmospheric Correction module and RPC (Rational Polynomial Coefficient) Orthorectification module was used for converting the original DN value to atmospheric surface reflectance, atmospheric correction and precise geometric correction of WFV images, respectively. Then, the Band Math module was used to calculate the vegetation indices of NDVI, NDWI and NDMI. The Co-occurrence measures module was used to extract image texture features of WFV images based on a sliding window with 3 × 3 pixels, and the texture indices mainly included Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment and Correlation. The DEM data were 90 m shuttle radar topography mission (SRTM) images (version V004) (http://srtm.csi.cgiar.org/ (accessed on 1 June 2018) in Geo-TIFF format. The Slope, topographic position index (TPI) and aspect were calculated based on the DEM. Then, all indices above mentioned were uniformly projected as UTM_Zone_47N (same as GF WFV).

Vegetation Community Classification and Accuracy Evaluation 2.3.1. Classification Method
The maximum likelihood estimate (MLE), NN, SVM and RF classification methods were employed. MLE assuming each statistic of different types in every band was normally distributed, the likelihood of each pixel belonging to a certain training sample was calculated. Finally, the type of pixel was determined based on the highest likelihood [36]. NN (also called artificial neural network, ANN) referred to a multi-layer network structure, the Levenberg-Marquardt function algorithm was selected for NN training. The number of neurons and hidden layers were determined based on a trial-and-error process [37]. SVM was constructed by a set of hyperplanes in high-or infinite-dimensional space, the higher the functional margin, the lower the generalization error of the classifier. The radial basis function (RBF) was used as the kernel function, and the optimal cost and gamma values were obtained for final classification [38,39]. The RF algorithm was constructed by the classification tree, which applied a set of decision trees to improve prediction accuracy. The bootstrap sample was employed to construct a decision tree. The training samples were constantly selected to minimize the sum of the squared residuals until a complete tree was formed. Multiple decision trees were formed, and voting was used to obtain the final prediction [40,41]. MLE, NN and SVM methods were performed in ENVI supervised classification toolboxes of Maximum Likelihood Classification, Neural Net Classification and Support Vector Machine Classification, respectively. RF method was performed in ENVI Extensions toolbox of Random Forest Classification [42].

Classification and Accuracy Evaluation
Given the classification accuracy and efficiency, three input datasets were used: (1) GF1/6 WFV spectral band (band1 to band8); (2) vegetation and texture indices (NDVI, NDWI, SAVI, Contrast, Correlation, Dissimilarity, Entropy, Homogeneity, Mean, Second moment and Variance); (3) vegetation, texture, and topography indices (DEM, Slope, Asp and TPI). About 70% of observation sites were selected randomly as a training set, and the rest were used to validate classification accuracy in each county. The standard confusion matrix was employed to evaluate the classification accuracy of images, and the overall accuracy (OA), Kappa coefficient (Kappa), user's accuracy (UA) and producer' accuracy (PA) based on the validation datasets were used to assess the precision of classification results.

Characteristics of Field Observation and Its Corresponding Multi-Indices
The distribution of observed sites was shown in Figure 1a. The vegetation communities of alpine meadow showed a considerable spatial heterogeneity. Among the 751 observed sites, the proportion of KH community is highest, with 56.32% of all observed sites. Followed by KP community (17.04% of all observed sites), the number of KP community observation sites were 68, 37 and 22 for Maqu, Zeku and Henan, respectively. The proportion of SM, MM, BS and Poaceae only accounted for 3.33~9.85% of all observed sites.
For the four types of alpine meadow grass communities and four types of land cover, the statistics of GF1/GF6 WFV image bands, vegetation indices, topography indices and texture indices were calculated in the study area (in Supplementary Materials). The result showed that the characteristics of multi-indices in alpine meadow vegetation communities were very similar, and it was difficult to distinguish with commonly used indices (Figure 4a,b,e). Even though eight land covers could be coarsely distinguished between each other in red edge bands (band5 and band6 of WFV image) and DEM, there was relatively large error in classification (with little difference in mean values and wide range in variation) (Figure 4c,d,f).

Accuracy Evaluation of the Different Classification Methods
Accuracy assessment of classification was performed with the validation samples listed in Table 4. Among the four classification methods, the RF method performed best, with the highest overall accuracy and Kappa coefficient ranged from 74.06 % to 83.92% and from 0.65 to 0.80 in three counties, respectively. This was followed by the SVM method, with an overall accuracy that ranged from 69.39% to 78.53% and Kappa coefficient that ranged from 0.60 to 0.73. The accuracies of the NN and MLE method were the worst (overall accuracy ranged from 40.78 % to 73.89%; Kappa coefficient ranged from 0.24 to 0.67). Among the three classifications input, in general, the MLE, SVM and RF methods based on the input data set of vegetation indices + texture + topography exhibited the best performance, followed by the spectrum and vegetation indices + texture. However, the performance of the NN method based on the above input data set showed contrary results to the MLE, SVM and RF methods.
The results of the standard confusion matrix were shown in Table 5. The PA and UA based on the RF method were highest in three counties, with 60.84 % to 97.23 % and 60.73 % to 78.09%, respectively. The PA and UA based on other methods showed lower value,

Accuracy Evaluation of the Different Classification Methods
Accuracy assessment of classification was performed with the validation samples listed in Table 4. Among the four classification methods, the RF method performed best, with the highest overall accuracy and Kappa coefficient ranged from 74.06% to 83.92% and from 0.65 to 0.80 in three counties, respectively. This was followed by the SVM method, with an overall accuracy that ranged from 69.39% to 78.53% and Kappa coefficient that ranged from 0.60 to 0.73. The accuracies of the NN and MLE method were the worst (overall accuracy ranged from 40.78% to 73.89%; Kappa coefficient ranged from 0.24 to 0.67). Among the three classifications input, in general, the MLE, SVM and RF methods based on the input data set of vegetation indices + texture + topography exhibited the best performance, followed by the spectrum and vegetation indices + texture. However, the performance of the NN method based on the above input data set showed contrary results to the MLE, SVM and RF methods. The results of the standard confusion matrix were shown in Table 5. The PA and UA based on the RF method were highest in three counties, with 60.84% to 97.23% and 60.73% to 78.09%, respectively. The PA and UA based on other methods showed lower value, and the classification results of the KP community were easily confused with other grass communities and land cover types.

Distribution and Area of KP Community
According to the vegetation community distribution map acquired by the RF method, the spatial distribution of the KP community was fragmented with large spatial heterogeneity and small area ( Figure 5). Among the three counties, the distribution of the KP community was mainly located in: the north, east and around the county urban area of Zeku County (around the town of Zequ, Qiakeri and Xipusha), with an area of 445.60 km 2 (6.82% of Zeku County); the northeast and central part of Henan County (east of county urban area, towns of Tuoyema and Duosun, and north of Saierlong), with an area of 176.76 km 2 (4.48% of Henan County); the part of county urban area, towns of Oulaxiuma, Muxihe and Awancang in Maqu County, with an area of 811.70 km 2 (8.59% of Maqu County). As a whole, the area of KP community reached 1434.07 km 2 , and accounted for 7.20% of the study area.

Distribution and Area of KP Community
According to the vegetation community distribution map acquired by the RF method, the spatial distribution of the KP community was fragmented with large spatial heterogeneity and small area ( Figure 5). Among the three counties, the distribution of the KP community was mainly located in: the north, east and around the county urban area of Zeku County (around the town of Zequ, Qiakeri and Xipusha), with an area of 445.60 km 2 (6.82% of Zeku County); the northeast and central part of Henan County (east of county urban area, towns of Tuoyema and Duosun, and north of Saierlong), with an area of 176.76 km 2 (4.48% of Henan County); the part of county urban area, towns of Oulaxiuma, Muxihe and Awancang in Maqu County, with an area of 811.70 km 2 (8.59% of Maqu County). As a whole, the area of KP community reached 1434.07 km 2 , and accounted for 7.20% of the study area.

Influence Factors of KP Community in the Qinghai-Tibet Plateau
Generally, the KP community builds almost closed, non-specific, golf-course like the lawn with a felty root mat. This characteristic mat not only protects soil against intensive trampling by herbivores, but also helps to cope with nutrient limitations enabling medium-term nutrient storage and increasing productivity and competitive ability of roots against leaching and other losses [43][44][45]. However, with browning (patchwise dieback of lawns), crack, collapse, fragmentation of KP community turf, the water budget [46], carbon cycle [47,48] and soil nutrition [44,45,49] have been significantly changed [10].
Pastoralism may have promoted the dominance of KP community and is a major driver for felty root mat formation [10]. However, the degradation of KP grassland may be caused by both human activities and climate change [9]. The mean annual precipitation in the northern and western parts of the QTP (the elevations ranged from 4400-4800 m) was less than 450 mm, with an increase of inter-annual variability towards the west [2,10]. Grassland suffered from co-limitation of summer rainfall and nutrient shortage [10,[50][51][52][53]. The types of grassland were diverse, but the species richness was low [10,15]. Hence, the ten distinct plant communities were described in this area [2]. The grassland is dominated by KP community in closed lawns with covers of 98%, and companion species less than 10 [10,43].
Our study area is located at the eastern edge of the QTP (including three counties), the mean elevation is 3758 m (Figure 1c) and mean annual precipitation ≥ 450 mm. The alpine meadow in study area consists of four types of vegetation communities, including (a) the Poaceae community (Elymus nutans + Stipa silena + Festuca ovina), (b) the Kobresia humilis community, (c) the Kobresia pygmaea community (KP), and (d) the denuded black soil ecosystem. Those communities consist of more than 40 species, with mosaics of KP community patches and grasses, other sedges and perennial forbs growing as rosettes and cushions [54,55]. Overgrazing is the main inducing factor for grassland vegetation community variation [5,10,56], but effect of climate still cannot be eliminated. Although we have mapped the distribution of KP community, the relative contributions from climatic and anthropogenic forces require further investigation. The main effect factor can be distinguished by combining the potential distribution based on the ecological niche model [57] and realistic distribution based on remote sensing, which is very important for alpine meadow protection.

Field Observation
KP community plays a vital role in alpine meadow degradation succession in QTP. However, its spatial distribution is difficult to map: on the one hand, the field observation data is lacking; on the other hand, the distribution of the KP community is under a dynamic variation with different disturbances [5,43]. The massive field observation is the basis of RS classification for grassland community. Traditional grassland vegetation community samples were obtained with the few field investigation, expert knowledge and literature reviews [11][12][13][14]58]. Field observation is mainly carried out at quadrat, plot and belt transection scales [15,59,60]. Due to the complex distribution of grassland vegetation communities, the field investigation is difficult and time-consuming. Meanwhile, the expert knowledge and literature reviews cannot meet the accuracy requirement of classification, because of the subjective bias, the dynamic climate and anthropogenic activities [15,61].
In this study, the field observation was performed by UAV based on FragMAP [14]. The resolution of each aerial photo is~0.87 cm and covers~35 × 26 m of ground at the height of 20 m, which is close to the traditional ground observation plot [25,62]. Moreover, the UAV is efficient and easy to operate (about 15 min to finish each observation site), which provides the possibility for rapid observation in large regions [24]. Most importantly, the waypoints, once established, can be repeatedly used (the error of two flights of the same waypoint is 1-2 m, and two photos on the same waypoint from two different flights are almost overlapped). It is suitable to monitor the dynamic variation of grassland communities in a long-term period [25,62].
Limited by the UAV control range and battery life, the size of ROI was only 250 × 250 m, and the proportion of image raster used for training classification is relatively small. Besides, most of field observation sites were located in the flat area, which was near major traffic roads. Therefore, the spatial distribution of KP community still had some uncertainty in other regions of the study area. Moreover, the vegetation communities were distinguished by manual visual interpretation, and it requires good knowledge of plant taxonomy and time-consuming. Hence, the automatic identification of vegetation community based on aerial photograph and deep learning algorithm requires further exploration.

Classification Variables
NDVI, NDWI and SAVI have been commonly used as the classification variables for grassland classification [20,33,59]. The vertical variation of grassland vegetation is significantly changed with topographic features in the QTP [63], hence, topographical factor is an important classification basis in alpine vegetation communities classification [64]. Additionally, texture features are also essential variables in object-based classification, which usually reflect local spatial information relating to the change of image tone [16,17]. The common method in texture feature extraction is the grey level co-occurrence matrix (GLCM). The texture metric includes angular second moment, contrast correlation, entropy, homogeneity, difference, average and standard degrees [18]. Incorporating texture feature information usually enhances the recognition of "the same object with different spectrum" or "the different object with same spectrum" [16][17][18].
Our results showed that, the threshold range of these RS indices for identifying the alpine meadow communities are commonly confused during extraction and identification. According to the descriptive statistical value of those RS indices corresponding to the four alpine meadow grass communities, the threshold range of KH was close to KP, and that of Poaceae was close to BS among the NDVI, NDWI and SAVI (Figure 6a-c). Although four grass communities could be distinguished in topography and texture metrics, there were relatively few differences and large errors (with little difference in mean values and wide range in variation) (Figure 6d-i). Therefore, it was difficult to distinguish the alpine meadow grass communities based on single variable and simple combinations [33][34][35]. RS classification accuracy can be improved by combining the RS, topographic and texture indices (Tables 3 and 4).
Due to large errors in spatial quantification of some variables (such as texture indices), the classification still has some limitations and uncertainties [29]. Hence, we consider using high spatiotemporal resolution images in future research, such as the Sentinel-2A/B satellite images, to reduce the effects of spatial heterogeneity on spectral reflectance and acquire more detailed texture features. Secondly, screening and reconstructing the remote sensing vegetation index: combining existing vegetation index, screening out indices that are more suitable for alpine meadow vegetation community classification.

Classification Method
Limited by the low temporal-spatial resolution, few spectrum band of RS images and field observations, most of natural grassland classification were applied in land use types (such as non-grassland, grassland, woodland, etc.) [33], different biophysics characteristics (for example, grassland with high, medium and low coverage) [34] and types with different climatic zones (e.g., groups and types of grassland) [35]. The most frequently used classification methods are visual interpretation, maximum likelihood classifiers, k-nearest neighbor and decision tree classification, and so on [65][66][67]. With the development of classification methods, the machine learning algorithm has obvious advantages in RS image classification [28,29]. However, the previous grassland classifications have been done at the vegetation type level, and few at the community level [8,9].

Classification Method
Limited by the low temporal-spatial resolution, few spectrum band of RS images and field observations, most of natural grassland classification were applied in land use types (such as non-grassland, grassland, woodland, etc.) [33], different biophysics characteristics (for example, grassland with high, medium and low coverage) [34] and types with different climatic zones (e.g., groups and types of grassland) [35]. The most frequently used classification methods are visual interpretation, maximum likelihood classifiers, k-nearest neighbor and decision tree classification, and so on [65][66][67]. With the development of classification methods, the machine learning algorithm has obvious advantages in RS image classification [28,29]. However, the previous grassland classifications have been done at the vegetation type level, and few at the community level [8,9].
Referenced with previously classification methods [20,58,[68][69], the ANN, AVM and RF were used to distinguish the alpine meadow grass communities based on RS, texture and topographic indices in the QTP. Our results demonstrated that the RF algorithm had higher overall accuracy than other algorithms by using the same training samples (with 74.06% to 83.92%). Compared with other methods, RF is a data-driven algorithm. With the increase of input dataset, classification accuracy is improved correspondingly [66,[70][71]. The RF algorithm can estimate complex nonlinear relationship and all the quantitative and qualitative information distributed within the models better; thus, these models are robust and fault-tolerant [69,70]. Moreover, the Referenced with previously classification methods [20,58,68,69], the ANN, AVM and RF were used to distinguish the alpine meadow grass communities based on RS, texture and topographic indices in the QTP. Our results demonstrated that the RF algorithm had higher overall accuracy than other algorithms by using the same training samples (with 74.06% to 83.92%). Compared with other methods, RF is a data-driven algorithm. With the increase of input dataset, classification accuracy is improved correspondingly [66,70,71]. The RF algorithm can estimate complex nonlinear relationship and all the quantitative and qualitative information distributed within the models better; thus, these models are robust and fault-tolerant [69,70]. Moreover, the input classification indices can be acquired by different multi-spectral remote sensing images, and it helps to integrate multi-source remote sensing data [66,70,71]. However, it is difficult to train the RF model effectively with a small sample dataset. RF algorithm composes a large sample decision tree, and classification is performed based on the voting results of each decision tree, thus, has a strong tolerance for data error [40,41]. Constructing decision trees consumes more time while performing random forest classification [70].

Conclusions
Based on the band spectral, vegetation indices, texture feature of GaoFen 1/6 wide field view images, topographic indices and UAV field observation, this study examined four classification methods and evaluated their accuracy. Our results showed that the characteristics of RS indices in alpine meadow vegetation communities were very similar, and it was difficult to distinguish the alpine meadow grass communities based on single variable or simple combinations. The KP community could be distinguished through the RF method based on combination of RS, texture and topographic indices. The spatial distribution of KP community was fragmented with large spatial heterogeneity and small area in three counties. The area was 1434.07 km 2 , which accounted for 7.20% of the whole study area. Our study demonstrated it was feasible to map at the community level using the satellite remote sensing, UAV surveying and machine learning methods. In future work, more detailed texture features derived from the high spatiotemporal resolution images are required to improve the grassland vegetation community classification.