A Semi-Automated Method to Extract Green and Non-Photosynthetic Vegetation Cover from RGB Images in Mixed Grasslands

Green (GV) and non-photosynthetic vegetation (NPV) cover are both important biophysical parameters for grassland research. The current methodology for cover estimation, including subjective visual estimation and digital image analysis, requires human intervention, lacks automation, batch processing capabilities and extraction accuracy. Therefore, this study proposed to develop a method to quantify both GV and standing dead matter (SDM) fraction cover from field-taken digital RGB images with semi-automated batch processing capabilities (i.e., written as a python script) for mixed grasslands with more complex background information including litter, moss, lichen, rocks and soil. The results show that the GV cover extracted by the method developed in this study is superior to that by subjective visual estimation based on the linear relation with normalized vegetation index (NDVI) calculated from field measured hyper-spectra (R2 = 0.846, p < 0.001 for GV cover estimated from RGB images; R2 = 0.711, p < 0.001 for subjective visual estimated GV cover). The results also show that the developed method has great potential to estimate SDM cover with limited effects of light colored understory components including litter, soil crust and bare soil. In addition, the results of this study indicate that subjective visual estimation tends to estimate higher cover for both GV and SDM compared to that estimated from RGB images.


Introduction
Fractional vegetation cover, defined as the percentage of vegetation that is vertically projected in a unit area [1], is an important indicator of plant growth [2], vegetation status [3], crop health [4], habitat selection [5] and ecosystem change [6]. Vegetation cover is also closely related to leaf area index, net primary productivity, biomass, soil stability, photosynthesis and ecological processes [6][7][8].
In grassland ecosystems, systematic, accurate and repeatable surveys of vegetation cover are essential for monitoring grassland condition, protection of soil erosion and grassland management [9][10][11].
Fractional cover estimation in grasslands relies on field measurements combined with remote sensing technology [7]. Field measured cover data are fundamental for quantitative models using remotely sensed images [12,13] and are necessary to validate empirical models to estimate grassland cover [6,14]. Visual (i.e., non-destructive) estimation is a commonly used field method for grassland vegetation cover [15]. This is a rapid and repeatable evaluation of vegetation cover [16] and is sufficiently accurate for relative (as opposed to absolute) assessments of cover data [7]. However, visual estimation

Field Data Collection
Fieldwork was performed during 20 June to 2 July 2014 in the peak growing season of GNP. A stratified random sampling design was used to select 14 sites with consideration of accessibility ( Figure 1: 4 sites in upland grassland, 5 in sloped grassland, 3 in valley grassland and 2 in disturbed communities). Two 100 m transects were surveyed perpendicular to one another and crossing in the center at each site. Twenty, 50 cm × 50 cm quadrats at 10 m intervals (excluding the center point) were surveyed along the transects.
This design is intended to collect the heterogeneity of biophysical parameters on the representative grasslands. Percent ground cover, including grass, shrub, forb, SDM, litter, moss, lichen, rock and bare soil coverage were visual estimated at each quadrat. The descriptive statistics of GV cover (i.e., sum of grass and forb cover), SDM and NPV cover (i.e., sum of SDM and litter cover) are shown in Table 2. Nadir (i.e., downward facing) RGB images were taken by a commercially available digital camera (Nikon S8000, Nikon Imaging Japan Inc., Tokyo, Japan) at each quadrat (i.e., the corresponding RGB pictures for each quadrat in the 14 sites of Table 2 are listed in Supplementary S2) at 1 m above the ground. A 0 • camera angle enables fractional cover estimation when compared to oblique angles tested [31]. Hyper-spectra (wavelength from 350 nm to 2500 nm) were also measured at each quadrat with an Analytical Spectral Devices (ASD) field-portable FieldSpec ® Pro Spectroradiometer between 10:00 a.m. and 14:00 p.m. under clear sky (i.e., without any cloud cover).

Methods
The methodological workflow for the proposed semi-automatic method included preprocessing digital RGB images, developing a python script to extract GV and SDM separately and calculating GV and SDM percentage cover automatically ( Figure 2). The result of semi-automatically estimated GV and SDM cover with visual estimated cover data and vegetation indices were validated based on hyperspectral remote sensing ( Figure 2). 2.3.1. Pre-Processing for the Field-Taken RGB Images RGB images were first cropped to the quadrat area and then processed to their actual size (50 cm × 50 cm) with 300 pixels/inch (dpi) using Adobe Photoshop CS6 ( Figure 2).
Because light conditions differed slightly among field-taken RGB images, blue, green and red bands of cropped pictures were standardized independently to maintain consistency among study sites (Equation (1)).
where is the standardized pixel value, is the original pixel value, is the mean value of all the pixel values in a single band, and ( ) is the standard deviation for all the pixel values 2.3.1. Pre-Processing for the Field-Taken RGB Images RGB images were first cropped to the quadrat area and then processed to their actual size (50 cm × 50 cm) with 300 pixels/inch (dpi) using Adobe Photoshop CS6 ( Figure 2). Because light conditions differed slightly among field-taken RGB images, blue, green and red bands of cropped pictures were standardized independently to maintain consistency among study sites (Equation (1)).
where DN std is the standardized pixel value, DN is the original pixel value, DN is the mean value of all the pixel values in a single band, and std(x) is the standard deviation for all the pixel values in a single band.
After each band for all the pictures was standardized, all the pixel values fit in a range from −1 to 1. Standardized images were then normalized (Equation (2)) as images with pixel value range from 0 to 1023 (10 bit integer data format).
where DN nor is the normalized pixel value, DN std is standardized pixel value, min(xDN std ) is the minimum value of all the pixel values in each standardized band, and max(DN std ) is the maximum of all the pixel values in each standardized band.

Developing a Python Script to Semi-Automate GV and SDM Cover Extraction from Preprocessed RGB Images
GV pixels in the RGB images were extracted based on the spectral characteristics of GV (i.e., reflectance for green leaves in the green band is larger than that in both red and blue bands; Equation (3)). GV pixels were masked out before further process for extracting SDM pixels.
where DN G nor , DN R nor , DN B nor are the normalized pixel value for the green band, red band and blue band of the field-taken RGB pictures, g 1 , g 2 are constants (i.e., default values were set to 60 in this study). The values of g 1 , g 2 were determined after exploring the spectral characteristics of green leaves for narrow leaved native grasses, shrubs, invasive species in disturbed communities (their values are discussed in Section 3.1).
Even though SDM has similar spectral characteristics such as litter, soil crust and bare soil, SDM in the canopy have much brighter color tong in all the three visible bands in field-taken RGB images because the understory components receive limited sunlight in comparison with the vegetation canopy. Therefore, SDM is extracted based on the criteria that the SDM has higher pixel values than the understory components in the normalized visible bands after GV pixels were removed (Equation (4)).

DN
where DN G nor , DN R nor , DN B nor are the normalized pixel values for green, red and blue bands of the field-taken RGB images, DN G nor , DN R nor , DN B nor are the mean pixel values of the green, red and blue normalized bands of the field-taken RGB images, d is a constant set as a default value of 1 for our study. The setting of constant d was discussed in Section 3.2 when litter and soil background with light color tong challenging the extraction of SDM.
After GV and SDM pixels were separated from the pre-processed RGB images, GV and SDM cover were calculated. Both GV and SDM pixel counts were divided by total pixel counts for fractional cover. These cover estimations were compared with field observed cover data by visual estimation.
All classification and cover estimation processes for both GV and SDM were conducted using a python script developed in this study (see the stand-alone python script and ArcToolbox in the Supplementary Materials and the description of ArcToolbox in Supplementary S1).

Validation of Extracted GV and SDM Cover from RGB Images
Alternative methods for GV and NPV cover estimation (i.e., vegetation indices based on hyperspectral remote sensing) were used to validate estimated GV and SDM cover from field-taken digital RGB images. Normalized difference vegetation index (NDVI), an index strongly correlated to GV, has been widely used to evaluate GV cover in grasslands [10]. The cellulose absorption index (CAI) is effective for estimating NPV fractional cover (i.e., including SDM and litter cover) from GV and soil background [32][33][34]. Therefore, NDVI calculated from field measured hyper-spectra (Equation (5)) was used to test the accuracy of GV cover extracted from field-taken RGB images based on linear regression analysis in R software (i.e., it is written by John Chambers and his colleagues of the Bell Laboratories, Murray Hill, NJ, USA). The CAI calculated from field-collected hyper-spectra (Equation (6)) was used to validate the accuracy of SDM extraction from field-taken RGB images with linear regression analysis in R software (i.e., the NPV cover used for linear regression with CAI is the sum of RGB image extracted SDM cover and visual estimated litter cover).
where ρ 800 and ρ 670 are the reflectance in the wavelength of 800 nm and 670 nm from field-collected hyper-spectra.
where ρ 2030 , ρ 2100 and ρ 2210 are the reflectance in the wavelength of 2030 nm, 2100 nm and 2210 nm from field-collected hyper-spectra.

Determination of the Constants g 1 , g 2 for GV Extraction
After data exploration of 25 sample RGB images for different species in different conditions ( Table 3, see the sample RGB photos in Supplementary S3), values of g 1 , g 2 were found to exceed 60 when the photograph was taken without high exposure under strong light (normal light conditions). Under normal light conditions, the minimum values of g 1 , g 2 for broad-leaved vegetation (e.g., Sweet Clover) was higher than narrow-leaved grasses (Table 3: id 5 and 6). When the RGB images were taken with high exposure or when the vegetation had begun to senesce, constant g 1 needed to be set at a value lower than 60 (32.09-44.13; Table 3). The value of g 2 was not affected by these conditions but was affected by bluish leaves (i.e., dominated western wheat grass and sagebrush). In these instances, the value of g 2 needed to be set lower (i.e., 32.09-40.12; Table 3).

Setting of the Constant d for SDM Extraction
SDM, as one part of grassland canopy, has higher pixel values in all three bands of field-taken RGB images than bare soil, litter and soil crust after GV pixels were masked out (Equation (4)). Therefore, it is effectively extracted by default constant d (set at 1) when the understory background, including soil, litter and soil crust, is dark (Figure 3a   Light colored, undecomposed litter has a large effect on the extraction of standing dead material (Figure 3e: visual estimation for standing dead materials is 40% and for litter is 20%) when the canopy cover (i.e., sum of GV and SDM cover) is low, thus, d needs to be set higher than the default value of one. The extracted SDM was 30.8% (Figure 3e1), 20.2% (Figure 3e2) and 6.5% (Figure 3e3) when d was set at 1, 1.2 and 1.5, respectively. Light colored soil crust as the canopy background also has a strong influence in SDM extraction in the study area (Figure 3f: visual estimation for standing dead materials is 15%). In this case, d must be set higher to reduce the influence of soil crust (i.e., moss and lichen). The extracted SDM was 27.8% (Figure 3f1), 8.6% (Figure 3f2) and 0.4% (Figure 3f3) when d was set at 1, 1.5 and 2, respectively. Light colored bare soil also influences the extraction of SDM with a default value of d (Figure 3g: visual estimation for standing dead material is 10%). d must be set higher to reduce the effects of bare soil (Figure 3g3: d was set at 2) providing high accuracy of standing dead material extraction (0.8%, Figure 3g3) compared to 11.9% extraction where some bare soil pixels were extracted (Figure 3g2: d was set at 1.5) and 25.9% extraction where a large amount of bare soil pixels were present in the extraction results (Figure 3g1: d was set at 1). Especially when canopy cover is low, light colored bare soil still has great effects on SDM extraction even though d is set appropriately for extracting SDM pixels (Figure 3h1: d was set at 1.5; Figure 3i1: d was set at 1.7).

GV and SDM Cover Estimated from RGB Images
Compared to subjective visual estimation, GV cover is under-estimated by the method developed in this study (Figure 4a). The difference between estimated GV cover and subjective visual estimated GV cover is larger when GV cover is lower (Figure 4a). Estimated cover of SDM in this study is underestimated compared to that from subjective visual estimation. The underestimation becomes more distinct when SDM cover is higher (Figure 4b).

GV and SDM Cover Estimated from RGB Images
Compared to subjective visual estimation, GV cover is under-estimated by the method developed in this study (Figure 4a). The difference between estimated GV cover and subjective visual estimated GV cover is larger when GV cover is lower (Figure 4a). Estimated cover of SDM in this study is under-estimated compared to that from subjective visual estimation. The underestimation becomes more distinct when SDM cover is higher (Figure 4b). Comparing estimated green and dead cover with field observed cover (the red solid line is the regression line; the blue line is 1:1). (a) Comparison between green vegetation (GV) cover from visual estimation (i.e., field observed green cover) and that from GV extraction by the method developed in this study. (b) Comparison between SDM from visual estimation (i.e., field observed standing dead cover) and that from SDM extraction from the method of this study.

Validation of GV and NPV Estimated from RGB Images
Based on the relationship between GV cover and NDVI, the estimated GV cover (Figure 5b: R 2 = 0.846, p < 0.001) in this study is more precise than that from subjective visual estimation (Figure 5a: R 2 = 0.711, p < 0.001). Comparing estimated green and dead cover with field observed cover (the red solid line is the regression line; the blue line is 1:1). (a) Comparison between green vegetation (GV) cover from visual estimation (i.e., field observed green cover) and that from GV extraction by the method developed in this study. (b) Comparison between SDM from visual estimation (i.e., field observed standing dead cover) and that from SDM extraction from the method of this study.

Validation of GV and NPV Estimated from RGB Images
Based on the relationship between GV cover and NDVI, the estimated GV cover (Figure 5b: R 2 = 0.846, p < 0.001) in this study is more precise than that from subjective visual estimation (Figure 5a: R 2 = 0.711, p < 0.001).
Theoretically, CAI can be used to evaluate NPV cover including cover of SDM and litter. There is no significant linear relationship of CAI and estimated SDM cover. Therefore, cover used for validation is comprised of total NPV including SDM and litter, thus, the estimated dead cover (Figure 6b) is the sum of the estimated cover of SDM and field observed litter cover. The R square of estimated dead cover increased from 0.687 (p < 0.001) with subjective visual estimation to 0.734 (p < 0.001) with the linear regression of CAI (Figure 6a). Theoretically, CAI can be used to evaluate NPV cover including cover of SDM and litter. There is no significant linear relationship of CAI and estimated SDM cover. Therefore, cover used for validation is comprised of total NPV including SDM and litter, thus, the estimated dead cover ( Figure  6b) is the sum of the estimated cover of SDM and field observed litter cover. The R square of estimated dead cover increased from 0.687 (p < 0.001) with subjective visual estimation to 0.734 (p < 0.001) with the linear regression of CAI (Figure 6a).

GV and SDM Cover Estimation Based on RGB Images and Visual Estimation
Our semi-automated method to classify RGB images predicts lower GV cover than visual estimates (Figure 4a). The difference between visual estimation and extraction of GV coverage from RGB images is higher when the GV coverage is relatively low (Figure 4a). Previous research also suggests that subjective visual estimation tends to predict higher cover (i.e., overestimation) than GV cover estimates from digital image analysis based on field-taken RGB pictures [8,35]. Macfarlane and Ogden's results show that subjective visual estimation accuracy is ±10-20% [16]. This indicates that green cover collected by visual estimation may consistently overestimate real GV cover. Moreover, the linear regression of GV cover and NDVI, an alternative method based on remote sensing imagery for GV cover estimation, indicates that GV cover extracted from field-taken RGB images (Figure 5b:  Theoretically, CAI can be used to evaluate NPV cover including cover of SDM and litter. There is no significant linear relationship of CAI and estimated SDM cover. Therefore, cover used for validation is comprised of total NPV including SDM and litter, thus, the estimated dead cover ( Figure  6b) is the sum of the estimated cover of SDM and field observed litter cover. The R square of estimated dead cover increased from 0.687 (p < 0.001) with subjective visual estimation to 0.734 (p < 0.001) with the linear regression of CAI (Figure 6a).

GV and SDM Cover Estimation Based on RGB Images and Visual Estimation
Our semi-automated method to classify RGB images predicts lower GV cover than visual estimates (Figure 4a). The difference between visual estimation and extraction of GV coverage from RGB images is higher when the GV coverage is relatively low (Figure 4a). Previous research also suggests that subjective visual estimation tends to predict higher cover (i.e., overestimation) than GV cover estimates from digital image analysis based on field-taken RGB pictures [8,35]. Macfarlane and Ogden's results show that subjective visual estimation accuracy is ±10-20% [16]. This indicates that green cover collected by visual estimation may consistently overestimate real GV cover. Moreover, the linear regression of GV cover and NDVI, an alternative method based on remote sensing imagery for GV cover estimation, indicates that GV cover extracted from field-taken RGB images (Figure 5b:

GV and SDM Cover Estimation Based on RGB Images and Visual Estimation
Our semi-automated method to classify RGB images predicts lower GV cover than visual estimates (Figure 4a). The difference between visual estimation and extraction of GV coverage from RGB images is higher when the GV coverage is relatively low (Figure 4a). Previous research also suggests that subjective visual estimation tends to predict higher cover (i.e., overestimation) than GV cover estimates from digital image analysis based on field-taken RGB pictures [8,35]. Macfarlane and Ogden's results show that subjective visual estimation accuracy is ±10-20% [16]. This indicates that green cover collected by visual estimation may consistently overestimate real GV cover. Moreover, the linear regression of GV cover and NDVI, an alternative method based on remote sensing imagery for GV cover estimation, indicates that GV cover extracted from field-taken RGB images (Figure 5b: R 2 = 0.846, p < 0.001) in this study is superior to the GV cover from the subjective visual estimation in the field (Figure 5a: R 2 = 0.711, p < 0.001). We also compared our GV extraction with the extraction results from Canopeo (http://www.canopeoapp.com), a powerful tool for measuring GV cover in grassland [36] which has been proven to show good performance for measuring GV cover of narrow-leaved vegetation [11]. The comparison results show that the extracted GV cover with our semi-automated method is consistent with that of Canopeo (R 2 = 0.86, p < 0.001) and GV cover by Canopeo also has high linear relationship with NDVI (R 2 = 0.85, p < 0.001). It indicates that the method developed in this study has high potential to assess GV cover effectively and accurately. Moreover, this method has batch capacity, which would effectively support field data collection.
The results of linear regression between NPV cover and CAI (an alternative method for NPV cover estimation based on remote sensing approaches) show that cover estimates from RGB images in this study (Figure 6b: R 2 = 0.734, p < 0.001) are superior to subjective visual estimates (Figure 6a: R 2 = 0.687, p < 0.001). Estimated cover of SDM based on field-taken RGB images in this study is lower than subjective visual estimates. The difference between visual estimation and extraction from RGB pictures for SDM cover becomes larger when SDM cover increases. When SDM cover is larger, SDM in the lower layer is darker than in the top layer, which challenges the extraction of SDM from field-taken RGB images. Therefore, SDM cover might be underestimated by RGB images when the cover of SDM is very high. However, SDM cover might be overestimated from RGB images when SDM cover is low with background soil that has a light tone.

Estimated Green Cover from RGB Pictures
After standardizing and normalizing red (band 1), green (band 2) and blue (band 3) from the RGB images (Equations (1) and (2)), we found that the green band had the highest pixel value for green vegetation. Therefore, the constants g 1 , g 2 (Equation (3)) can be used as thresholds to separate green vegetation from SDM, litter, soil crust (moss and lichen), rocks and bare soil. In this study, green vegetation was extracted accurately in most cases when g 1 , g 2 were set to the default value of 60 (Figure 7a-d).
However, default values of g 1 , g 2 should be tied to vegetation type, phenology stage, soil crust and the light conditions when taking RGB pictures. Previous research indicates that GV cover extracted from field-taken RGB images is influenced by resolution, exposure and ground complexity [3]. In our study area, sagebrush and western wheatgrass are a pale blue color (Figure 7e,f). For this case, we adjusted g 2 lower (32) and moved g 1 to be lower than 60 to capture more sage leaves (Figure 7e,f; g 1 , g 2 were set as 40 and 32, respectively). When soil crust, especially moss with green color, influences green vegetation extraction from RGB pictures, g 1 could be set higher than g 2 because the difference between normalized green band and normalized blue band is far greater than the difference between normalized green band and normalized red band for green moss (Figure 7g; g 1 , g 2 were set as 60 and 40, respectively). Because the python script is designed to extract GV and SDM pixels separately to quantify ground cover of GV and SDM (Figure 7h; g 1 , g 2 were set as 30 and 60, respectively). If the vegetation is in early or late senescence, g 1 should be set as a lower value than the default 60 to extract more GV pixels which are not completely senesced.
When RGB photographs were taken near noon, the issue of high exposure reduced the greenness in the green band. Therefore, g 1 should be set lower than 60 (32, Table 3). Even though it improves the accuracy to extract green vegetation from RGB images by changing the parameters g 1 , g 2 in the python script we developed, the effects of sage and green moss are still not eliminated. They are, however, reduced. Green vegetation is effectively extracted by our method when GV was overlapping in the original RGB images (Figure 7) and compared with the GV extraction from Canopeo. When green moss was present, this principle did not necessarily hold (Figure 7g). For all these limitations of GV cover estimation, we suggest taking RGB pictures of each quadrat in the maximum growing season to avoid the senesced vegetation issue and avoid high exposure issue at noon. When RGB photographs were taken near noon, the issue of high exposure reduced the greenness in the green band. Therefore, should be set lower than 60 (32, Table 3). Even though it improves the accuracy to extract green vegetation from RGB images by changing the parameters , in the python script we developed, the effects of sage and green moss are still not eliminated. They are, however, reduced. Green vegetation is effectively extracted by our method when GV was overlapping in the original RGB images (Figure 7) and compared with the GV extraction from Canopeo. When green moss was present, this principle did not necessarily hold (Figure 7g). For all these limitations of GV cover estimation, we suggest taking RGB pictures of each quadrat in the maximum growing season to avoid the senesced vegetation issue and avoid high exposure issue at noon.

Estimated SDM from RGB Images
SDM has high pixel values in all three normalized visible bands. Equation (4) was designed based on this concept. Green vegetation was masked out for the normalized RGB images before extracting SDM to completely eliminate effects from green canopy, but the averaged pixel values of each normalized band were calculated before the green cover was masked out (Equation (4)). The parameter d is designed to separate standing dead cover from litter and a light soil background (Equation (4)). In this way, SDM was extracted accurately under moderate and high canopy cover (Figure 3a,b; d was set as default value 1).
Undecomposed litter has similar spectral characteristics as SDM. When the canopy cover (sum of GV and SDM cover) is low, litter has large effects on the accuracy for extracting SDM pixels. In normalized RGB images, litter is slightly darker than SDM in all the three bands because light exposure differs for the canopy and understory, and the color tone of litter becomes darker when it begins to decompose. To reduce the impact of litter when extracting SDM, the constant d was set higher (in the range one to two in this study) than the default value of one (Figure 3e3; d was set as 1.5).
Dry bare soil with light color tone is another issue for extracting SDM when the canopy cover is low. The errors caused by light soil background can be reduced by setting the constant d to a higher value (Figure 3g3; d was set as two). When the actual SDM cover is high, SDM cover may be underestimated (Figure 3c,d) with our method. SDM in the lower canopy has a darker color tone than that in the upper canopy when the SDM cover is high. Thus, the SDM in the lower canopy will be treated as litter to be excluded in the output of SDM cover (Figure 3c1,d1). In this specific case, we set the parameter d lower than the default value 1 (Figure 3c1,d1 when d was set as default value 1; Figure 3c2,d2 when d was set as 0.7; Figure 3c3,d3 when d was set as 0.5). In addition, the extracted results of SDM were overestimated due to the influence of soil crust covering the ground surface (Figure 3f; Figure 3f1 with d set as one). The effects of soil crust were reduced when d was set to 1.5 ( Figure 3f2). However, cover was underestimated when d was set to two to eliminate the effects of soil crust (Figure 3f3). each normalized band were calculated before the green cover was masked out (Equation (4)). The parameter is designed to separate standing dead cover from litter and a light soil background (Equation (4)). In this way, SDM was extracted accurately under moderate and high canopy cover (Figure 3a,b; was set as default value 1). Undecomposed litter has similar spectral characteristics as SDM. When the canopy cover (sum of GV and SDM cover) is low, litter has large effects on the accuracy for extracting SDM pixels. In normalized RGB images, litter is slightly darker than SDM in all the three bands because light exposure differs for the canopy and understory, and the color tone of litter becomes darker when it begins to decompose. To reduce the impact of litter when extracting SDM, the constant was set higher (in the range one to two in this study) than the default value of one (Figure 3e3; was set as 1.5).
Dry bare soil with light color tone is another issue for extracting SDM when the canopy cover is low. The errors caused by light soil background can be reduced by setting the constant to a higher value (Figure 3g3; was set as two). When the actual SDM cover is high, SDM cover may be underestimated (Figure 3c,d) with our method. SDM in the lower canopy has a darker color tone than that in the upper canopy when the SDM cover is high. Thus, the SDM in the lower canopy will be treated as litter to be excluded in the output of SDM cover (Figure 3c1,d1). In this specific case, we set the parameter lower than the default value 1 (Figure 3c1,d1 when was set as default value 1; Figure 3c2,d2 when was set as 0.7; Figure 3c3,d3 when was set as 0.5). In addition, the extracted results of SDM were overestimated due to the influence of soil crust covering the ground surface (Figure 3f; Figure 3f1 with set as one). The effects of soil crust were reduced when was set to 1.5 ( Figure 3f2). However, cover was underestimated when was set to two to eliminate the effects of soil crust (Figure 3f3). Flowers, especially white flowers in the mixed grassland, have large effects on extracting SDM (Figure 3f, Figure 8a

Conclusions
Our main conclusions are: (1) based on the linear relationship with NDVI, GV cover extracted from the method developed in this study (R 2 = 0.846, p < 0.001) is superior to that from subjective visual estimation in the field (R 2 = 0.711, p < 0.001), and the extracted GV cover is consistent with that

Conclusions
Our main conclusions are: (1) based on the linear relationship with NDVI, GV cover extracted from the method developed in this study (R 2 = 0.846, p < 0.001) is superior to that from subjective visual estimation in the field (R 2 = 0.711, p < 0.001), and the extracted GV cover is consistent with that estimated by Canopeo (i.e., a powerful tool for measuring GV cover in grassland). (2) The semiautomatic method of this study has high potential to extract SDM cover when the canopy cover (including both GV and SDM cover) is high or when the understory including the effects of litter, soil crust and bare soil, is limited. (3) Subjective visual estimation in the field tended to predict higher cover for both GV and SDM compared to that estimated from RGB images in this study.
Supplementary Materials: The following are available online at http://www.mdpi.com/1424-8220/20/23/6870/s1, Supplementary S1: explanation of the parameters in the python script for extracting green cover and standing dead cover from field-taken RGB images; Supplementary S2: RGB pictures for each quadrat in 14 sites; Supplementary S3: Sample RGB pictures for data exploration of constants g 1 , g 2 in Table 3; Python script, ArcToolbox.
Author Contributions: D.X. developed the initial idea, conducted fieldwork, developed the methodology and the python script, analyzed the data and wrote the manuscript. Y.P. improved the methodology and the python script. X.G. guided the organization of the initial idea and research direction, and revised the manuscript. All authors have read and agreed to the published version of the manuscript.