Aboveground Biomass Estimation and Time Series Analyses in Mongolian Grasslands Utilizing PlanetScope Imagery

Abstract

Mongolia, situated in central Asia and bordered by Russia to the north and China to the south, experiences a semi-arid climate across most of its territory. Grasslands are pivotal in Mongolia’s agricultural sustainability and food security, facing rapid changes in the last two decades that underscore the ongoing need for innovative approaches to assess vegetation conditions. This study aims to evaluate grassland biomass measurement and prediction through the analysis of high-resolution satellite data. By conducting a time series assessment of grazing-induced changes in vegetation dynamics at the long-term monitoring sites of the Botanic Garden and Research Institute, Mongolian Academy of Sciences, we seek to refine our understanding. The investigation covers biomass estimation across various Mongolian grassland landscapes, encompassing desert, steppe, and mountain regions. Spanning the grassland growing season from May 2020 to October 2023, the research leveraged diverse ground data types, including surface reflectance measurements, geographic coordinates for satellite data correction, and aboveground dry biomass. These components were instrumental in developing a biomass estimation model reliant on establishing correlations between the satellite-derived Normalized Difference Vegetation Index and biomass. The predicted biomass facilitated the time series map analysis and dynamic analysis. The PlanetScope surface reflectance correlates strongly at 0.97 with field measurements, indicating robust relations. Biomass and the Normalized Difference Vegetation Index show correlations of 0.82 for dry grassland, 0.80 for mountain grassland, and 0.65 for desert grassland, with a combined correlation coefficient of 0.62, revealing distinct characteristics across these grasslands. Time series dynamic analysis reveals rising biomass differences between grazed and ungrazed areas, suggesting potential grassland degradation. Variations in the slope coefficient of biomass differences among grassland types indicate differing degradation patterns, emphasizing the need for effective grazing management practices to sustain and conserve Mongolian grasslands. This highlights the potential of remote sensing in monitoring and managing grassland ecosystems.

1. Introduction

Spanning millions of years, grasslands cover about 38% of the Earth’s terrestrial land, excluding Greenland and Antarctica [1,2]. Predominantly situated in temperate regions, these ecosystems have undergone extensive modifications and stand among the most threatened terrestrial ecosystems globally [3].

Mongolia boasts a vast landscape of 156 million hectares, encompassing both mountains and extensive grasslands covering 112 million hectares. Within these grasslands, 110.3 million hectares serve as pasture, while an additional 1.7 million hectares are allocated for agriculture, constituting roughly 71.8% of the country’s total area, as of 2022 [4]. Supporting a population of 3.4 million across 941 K households, approximately 26% of these households rely on traditional animal husbandry for their livelihoods, contributing to a total count of 71 million livestock, including sheep, goats, cattle, horses, and camels [5]. As a result, grasslands play a pivotal role in livestock production, the economy, and tourism value due to their natural beauty, providing livelihoods for numerous individuals [6].

In recent years, global grasslands have faced significant degradation due to a combination of human activities and extreme climate events [7]. Within Mongolia, estimates from a national report evaluating grassland health indicate that roughly 58% of the country’s grassland areas are now classified as degraded [8]. Overgrazing stands as a pivotal factor contributing to grassland degradation in Mongolia [8]. Hence, the imperative to enhance effective grassland management and monitoring is becoming increasingly essential, urging the quest for innovative solutions.

Presently, grassland monitoring in Mongolia relies on ground station data utilizing traditional methods [8] heavily reliant on human intervention. This approach faces challenges related to extensive coverage, measurement frequency, and high costs. Remote sensing technology has significantly enhanced grassland monitoring operations, swiftly supplanting conventional approaches. It presents notable advantages in convenience, efficiency, and cost reduction [9]. Studying grasslands is a complex process and remote sensing provides an array of parameters for monitoring purposes, encompassing above-ground biomass (AGB), primary productivity, fractional vegetation cover (FVC), leaf area index (LAI), etc. [6,7,10]. Biomass serves as a crucial indicator in global carbon cycling, demonstrating grasslands’ carbon sequestration potential and offering direct insights into ecosystem health, facilitating overgrazing monitoring [11,12].

When modeling biomass estimation based on satellite remote sensing, both satellite data and ground data parameters are utilized. However, the ground field data often comes in small sizes, typically around 50 cm × 50 cm to 1 m × 1 m, posing challenges in covering larger areas such as 10 m × 10 m to 100 m × 100 m. To overcome this limitation, the ground data needs to be converted to the required size before being integrated into the estimation modeling process. Mongolian grasslands exhibit non-uniform grass growth due to livestock grazing activities. This variability can potentially affect the accuracy of biomass estimation models when using converted ground data, leading to a potential decrease in accuracy when applied to Mongolian grasslands. Until now, the highest resolution satellites employed for Mongolian grasslands encompass Sentinel-2 and Landsat 8 with 10 m and 30 m pixel resolutions, respectively [13,14]. Furthermore, in these conducted studies, ground field measured data were converted and spectral measurements for satellite data collection in the research area were not made. Therefore, it is important to ensure the utilization of corrected higher-resolution satellite data with pixels that approximate the size of the ground field area, along with untransformed ground measurement data, for estimation modeling.

The purpose of this study is to assess grassland aboveground dry biomass (ADB) measurement and prediction based on high-resolution satellite data by performing a time series analysis of grazing vegetation dynamic changes at the long-term monitoring sites of the Botanic Garden and Research Institute, Mongolian Academy of Sciences (BGRI-MAS). Research for biomass estimation has been carried out in diverse grassland terrains, including the desert, steppe, and mountain areas of Mongolia.

The time series has been analyzed for both grazed and ungrazed areas to determine a quantitative estimation of the grassland degradation characteristics. It is a crucial indicator for the grassland ecosystem of Mongolia, as well as the Central Asian region. Furthermore, this study innovates by being the first to integrate high-resolution satellite data from PlanetScope with ground measurement corrections in Mongolian grasslands.

2. Methodology

This study aimed to investigate the potential of high-resolution PlanetScope satellite imagery for estimating ADB in the Mongolian grassland area. The method relied on establishing a correlation between the satellite-derived Normalized Difference Vegetation Index (NDVI) [15] and ADB measured in the field. The research was conducted in three distinct grassland areas: desert grassland, dry (steppe) grassland, and mountain grassland.

The study process consisted of three main parts: data collection and processing, modeling, and analysis (Figure 1). The first section is divided into ground data and satellite data. Ground data included measurements of biomass, surface reflectance, and geometric coordinates of both the measurement points and ground control points. These on-site measurements were conducted with precision by our dedicated research team.

The biomass collected using the hand harvesting method in the specific field area of the BGRI-MAS includes both ungrazed and grazed areas. The BGRI-MAS has been collecting biomass data annually since 2009. In the field, wet biomass is gathered along with litter. Initially, plants and litter are classified separately, after which the wet plants are dried in the laboratory to obtain dry biomass.

The satellite data chosen for this study is PlanetScope, which offers high-temporal (daily) and high-spatial (3–3.7 m) resolution. We have collected PlanetScope Ortho Tile and Scene images that are free from cloud cover, and these images have already undergone geometric and radiometric correction. However, we conducted additional corrections specific to Mongolian grassland based on our ground data and methods. Geometric registration for the satellite data was conducted using the ground control point method, which involved utilizing a prepared white square sheet, and ground measurements of surface reflectance were performed for the radiometric correction of the red and near-infrared (NIR) bands of PlanetScope Dove-R (PS2.SD) and SuperDove (PSB.SD) sensors.

After obtaining the prepared and corrected data, we utilized the linear regression method to develop the biomass estimation model. We opted for the simple linear regression method, which relies solely on satellite data and is independent of additional parameters, making it suitable for assessing the accuracy of high-resolution PlanetScope images for biomass estimation in Mongolian grasslands.

Ultimately, time series mapping and dynamic analysis of ADB were conducted using the predicted ADB for each type of grassland. The study area includes both grazed and ungrazed (fenced) areas, providing opportunities for comparison and indicating grassland changes under grazing.

3. Study Area

The study was conducted in Mongolia, a country located between 41°34′ to 52°09′N latitude and 87°45′ to 119°56′E longitude, covering a total area of 1,562,950 km².

Mongolia’s climate zones comprise Subarctic, Arid, and Semiarid regions, featuring five distinct grassland types: high mountain, mountain, dry (typical), meadow, and desert [16]. The survey focused on three main grassland areas: desert grassland (25.41% of Mongolia), dry grassland (22.05% of Mongolia), and mountain grassland (12.99% of Mongolia), collectively covering approximately 60% of Mongolia’s land area and constituting 83% of the total Mongolian grassland area [17]. The ground survey was carried out at the long-term vegetation monitoring sites of the BGRI-MAS. These sites are located in desert, dry, and mountain grassland, as described in Figure 2.

The BGRI-MAS, a renowned institute with a 50 year history, serves as the primary research institution for vegetation and soil studies in Mongolia. They have established long-term vegetation monitoring sites with ungrazed areas (fences) covering all seasons, measuring 1 ha in the dry grassland and mountain grassland in 2009, and 0.5 ha in the desert grassland in 2010 (Figure 3). This setup enables a wide range of studies. For instance, comparing ungrazed and grazed areas helps identify the impact of external activities on grasslands. These sites’ locations were carefully chosen based on extensive soil and vegetation research to represent their respective regions accurately. Furthermore, comprehensive field measurements of vegetation biodiversity, biomass, cover, classification, and soil characteristics are conducted annually in these areas. Therefore, these regions were selected for the study due to their representativeness and the availability of infrastructure for reliable fieldwork.

The desert grassland has low plant coverage, receiving an annual mean precipitation between 120 and 200 mm, an annual mean temperature of 0.6 to 5.3 °C, and the dominant soil type is Gobi brown. The dry grassland features moderate plant coverage, experiencing an annual mean precipitation from 200 to 270 mm, an annual mean temperature between −2.6 and 1.2 °C, and the predominant soil type is chestnut. In contrast, the mountain grassland exhibits abundant vegetation, with an annual mean precipitation ranging between 130 and 370 mm, an annual mean temperature from −4.3 to −3.6 °C, and the primary soil type is dark chestnut.

The desert grassland area has less vegetation cover compared to the fenced and unfenced areas, as depicted in Figure 3a. The fenced area displays more grass than the unfenced area, as shown in Figure 3d,e. Figure 3b displays significant differences in grass cover between the dry grassland in both the fenced and unfenced areas. The top view of grass growth is evident in Figure 3f,g. The mountain grassland has a higher density than other grassland areas. From a side view, it looks similar to the fenced and unfenced areas in Figure 3c. However, when viewed from the top, it exhibits different growth patterns as shown in Figure 3h,i.

4. Data Collection and Processing

4.1. Aboveground Dry Biomass

The aboveground dry biomass with geographic coordinates was measured by the BGRI-MAS team from 2020 to 2022 and was jointly measured by the Yamaguchi University (YU) team and the BGRI-MAS team in 2023.

For the collected aboveground biomass, a total of 85 plots (

Table 1. Biomass measurement numbers by year.

	Measurement Year	2020	2021	2022	2023	Biomass Measurement /2020–2023/
Grassland Type
Desert grassland		7	6	8	12	33
Dry grassland		2	6	8	10	26
Mountain grassland		0	8	7	11	26
Total		13	20	23	33	85

) were utilized, employing the hand harvesting method at the base of the plants in three different study areas from July 25 to August 10, corresponding to the peak period of plant growth. Hand harvesting over large areas is challenging and the plots were selected from large homogeneous areas independent of the PlanetScope imagery resolution. Therefore, we opted for a plot size of 1 m × 1 m. The locations of the measurements were recorded using a handheld GPS Garmin Etrex 20×.

In the field area, all vegetation is harvested together due to the difficulty in distinguishing between green live plants and litter. To accurately separate green live plants from litter, this process is meticulously conducted in the laboratory. Additionally, the wet biomass is measured before drying, as the moisture content of the plants impacts their weight. Subsequently, dry biomass is obtained after the drying process. Therefore, dry biomass was utilized.

The drying process was conducted in the laboratory of the BGRI-MAS for 72 h at a temperature of 70 °C (Figure 4c). Then, the dried biomass was weighed, and aboveground dry biomass values were expressed as grams per square meter (g/m²).

4.2. Satellite Data

Within the study, the analysis integrated 343 high-resolution PlanetScope satellite images including 136 for mountain grassland, 107 for dry grassland, and 100 for desert grassland.

The PlanetScope constellation, consisting of approximately 130 satellites orbiting in low-Earth, sun-synchronous paths at altitudes between 450 and 580 km [18], possesses the ability to capture near-vertical images of the Earth’s entire land surface daily. Each satellite, designed as a 3U CubeSat measuring 10 cm × 10 cm × 10 cm, captures multiband images at a spatial resolution ranging from 3.123 to 4.1 m with a 12-bit radiometric resolution. Presently, the PlanetScope constellation operates three generations of satellites: Dove or PS2, Dove-R or PS2.SD, and Super-Dove or PSB.SD, respectively. Among the instruments offered by Planet, this study exclusively utilized data from the PS2.SD and PSB.SD instruments due to the similarity in wavelengths among the blue, green, red, and near-infrared bands. PlanetScope images may contain four or eight spectral bands, depending on the instruments used, as detailed in

Table 2. PlanetScope constellation and sensor specification.

Instrument	PS2.SD	PSB.SD
Spectral Bands	Blue: 464–517 nm Green: 547–585 nm Red: 650–682 nm NIR: 846–888 nm	Coastal Blue: 431–452 nm Blue: 465–515 nm Green I: 513–549 nm Green II: 547–583 nm Yellow: 600–620 nm Red: 650–680 nm Red-Edge: 697–713 nm NIR: 845–885 nm
Spatial Resolution	3.125 m	3.125 m
Revisit time	Daily at nadir	Daily at nadir

.

The research focused on selecting PlanetScope Ortho Tile and Scene images, emphasizing bottom-of-atmosphere (BOA) reflectance products designed as Level 3B by Planet, which underwent radiometric and sensor correction [18]. The atmospheric correction process employed the 6 s radiative transfer code [19]. All pixel values are stored in 16-bit Digital Number Tag Image File Format (TIFF). These exclusively PlanetScope images were acquired through the Planet Explorer website “https://www.planet.com/explorer/ (accessed on 29 December 2023)”.

4.3. Geographic Coordinates and Satellite Data Geometric Registration

Geographic coordinate measurements are indispensable when working with high-resolution (3 m) satellite data, particularly given the small ground sampling area (1 m × 1 m for biomass and 3 m × 3 m for surface reflectance). Ensuring the accurate matching of the geographic coordinates of ground sampling points with those of satellite image coordinates is vital for facilitating precise comparisons between ground and satellite images, thus significantly improving analysis accuracy. To achieve this, geometric registration was conducted for satellite images using the ground control point method, and the geographical coordinates of all measurement points were recorded.

During the study conducted across three sites (Figure 2) from 20 July to 1 August 2023, the geographic locations of all 149 measurements were meticulously recorded using handheld GPS Garmin Etrex 20× instruments (Figure 5a). These location measurements include three types of geographic coordinates: the first for the geo-location of biomass and surface reflectance measurement plots, the second for marking the geographic coordinates of fence corners and white sheet corners, and the third for measuring the distance between the measurement points.

For example, in Figure 5b,c, the biomass measurement point is marked in red, while orange, yellow, and green points present the satellite image geometric registration points.

The satellite image geometric registration process was carried out using the ground control point (GCP) method within the ENVI 5.6.2 software. In the grassland, challenges in locating ground control points led to the adoption of innovative methods in 2023, such as using white sheets sized 10 m × 10 m and 3 m × 3 m. In preceding years, fences and nearby objects like buildings within the research area served as ground control points. Geometric registration was performed on a total of 343 PlanetScope images during the study.

For instance, Figure 6 illustrates the geometric registration for a PlanetScope 3B-scene image captured on 1 August 2023, over the study site of the dry grassland area. Figure 6a presents the field view, Figure 6b displays the PlanetScope image before geometric registration, and Figure 6c showcases the PlanetScope image after geometric registration. During this registration process, red points served as ground control points, with white sheets specifically utilized for this purpose, resulting in a coordinate difference of 4.2 m. Similar processes are depicted in Figure 7 for desert grassland and Figure 8 for mountain grassland, revealing coordinate differences between ground measurements and satellite images of 3.3 m and 6 m, respectively.

4.4. Surface Reflectance and Satellite Data Radiometric Correction

The PlanetScope level 3B images are already radiometrically corrected, and the bottom-of-atmosphere surface reflectance image products [18] have been utilized in this study. However, we conducted additional radiometric correction by measuring surface reflectance in the research field area to enhance the accuracy assessment.

Spectral reflectance measurements were conducted at the three study sites (Figure 2) from 26 July to 1 August 2023, simultaneously with biomass data collection. An ASD FieldSpec HandHeld 2 Spectroradiometer (Figure 9a) was utilized for measurements, with a wavelength range of 325 nm to 1025 nm and a field of view angle of 25 degrees [19].

The measurement design involved setting the measurement height ‘h’ at 1.2 m and calculating the ground field area ‘S’ as 0.23 m² (Figure 9b). In each plot, 16 measurements were conducted, and the average values of these measurements were used for correlation. The plot size of 3 m × 3 m closely matched the resolution of PlanetScope imagery (Figure 9c). All measurements were taken between 10:00 AM and 12:00 PM to align with the timing of PlanetScope satellite image capture. Measurements of a white reference panel (Spectralon plate, Labsphere Inc. North Sutton, NH, USA) were taken immediately before each spectral reading. A total of 384 spectral measurements were obtained from 24 plot fields across the three study sites (Figure 9d).

PlanetScope PS2.SD and PSB.SD instrument images were used for this study, and their wavelength ranges are nearly identical (Table 2). Therefore, the same correction estimation method was applied to both of these instrument images.

The radiometric correction was carried out in two stages. First, all ground-measured surface reflectance was calculated for the red and NIR band wavelength range of PlanetScope PS2.SD and PSB.SD instruments. Second, the PlanetScope surface reflectance value was calculated from a 16-bit GeoTiff image by scaling 10,000 to a digital number [20]. All images were then converted to surface reflectance from the digital number using Equation (1).

$${SR}_{planet}={DN}_{planet}\times Scale$$

(1)

where

SR_planet—Surface reflectance value of PlanetScope image

DN_planet—Digital number of PlanetScope image

Scale—Scale factor from GeoTIFF image metadata. Equal to 0.0001.

After the conversation, a linear regression correlation method was applied to establish a correlation between ground surface reflectance and PlanetScope surface reflectance for separate red and NIR bands. In this study, the NDVI was derived for AGB estimation. Therefore, corrections were conducted for red and near-infrared bands.

The resulting correlations for the grassland study area are shown in Figure 10, and the red band has a correlation of R² = 0.9788 and the NIR band of R² = 0.9728, respectively.

These results indicate a strong relationship between PlanetScope surface reflectance and ground-measured surface reflectance, providing support for using Equation (2) for the red band and Equation (3) for the NIR band of the PlanetScope PS2.SD SDB.SD instrument’s image.

y = 2.1008x + 0.0881

(2)

y = 2.0886x + 0.13

(3)

The correlation results from 2023, Equations (2) and (3), were also applied to additional three year PlanetScope images (2020–2022).

5. Model Development and Time Series Analyses

5.1. Model Development

Accurate biomass estimation is crucial and has been extensively explored. Biomass estimation models typically fall into two categories [7]: parametric models (e.g., linear, logarithmic, exponential, and other regression-based methods) [21,22,23] and non-parametric models (such as support vector machine (SVM), random forest (RF), and artificial neural networks (ANNs), which are part of machine learning) [24,25,26,27].

One of our study objectives is to evaluate the accuracy of high-resolution PlanetScope imagery for biomass estimation in Mongolian grasslands. For this purpose, we chose the simple linear regression method because it relies solely on satellite data and is independent of additional parameters. This method establishes a relationship between aboveground dry biomass and the corrected PlanetScope NDVI. Specifically, the model utilizes a simple linear regression with only one regressor variable, as is characteristic of such models [28].

The simple linear regression model is designed using Equation (4) [28].

$$Y={\beta }_0+{\beta }_{1}X$$

(4)

where $Y$ is the dependent variable, $X$ is the independent variable, ${\beta }_0$ is the intercept coefficient, and ${\beta }_1$ is the slope coefficient.

The NDVI is calculated using Equation (5) [15,29].

$${NDVI}_{PlanetScope}=\frac{{NIR}_{PlanetScope}-{Red}_{PlanetScope}}{{NIR}_{PlanetScope}+{Red}_{PlanetScope}}$$

(5)

For modeling purposes, we utilized 85 ground field-measured datasets (see Table 1) in conjunction with corrected satellite data, as detailed in the data collection and processing section.

The model’s quality is assessed using the widely employed statistical coefficient of determination as described in Equation (6a) and the root-mean-square error (RMSE) outlined in Equation (6b) [28,30].

$$R^2=1-\frac{{\sum }_{i=1}^n{\left(y_i-\widehat{y}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}$$

(6a)

where $i$ is the number of observations, $y_i$ is the estimated variable, $\widehat{y}$ is the average of the estimated variables, and $\overline{y}$ is the average of the predicted variables.

$$RMSE=\sqrt[]{\frac{1}{n}\sum _{i=1}^n{(y_i-y_{obs})}^2}$$

(6b)

where $i$ is the number of observations and $y_{obs}$ is the observed variable.

5.2. Time Series Analyses

Time series analyses were conducted for the last four years, from 2020 to 2023, using predicted ADB derived from PlanetScope satellite data.

Initially, time series mapping analysis was performed to create predicted ADB maps for the same day of each year.

Secondly, the time series grazed and ungrazed ADB dynamic analysis focused on grazed and ungrazed grassland areas during the vegetation growing season over the past four years, from May 2020 to September 2023. The ungrazed area is fenced off to protect it from animals and other human factors, facilitating comparative analysis with the grazed area. Another advantage is that grazed and ungrazed study areas are situated adjacent to each other, experiencing the same climate and weather conditions.

A total of 343 PlanetScope images were utilized for the analysis of three grassland areas. In this time series analysis, average values were calculated separately for four sampling points in ungrazed areas (orange points) and grazed areas (red points) across all images, as shown in Figure 11.

The average ungrazed biomass was calculated using Equation (7).

$$\mathrm{Ungrazed}\_\mathrm{Biomass}=\frac{{\sum }_1^4{Ungrazed}_{point(1-4)}}{4}$$

(7)

where${Ungrazed}_{point(1-4)}$ are the ADB values for the four ungrazed sample points.

The average grazed AGB was calculated using Equation (8).

$$\mathrm{Grazed}\_\mathrm{Biomass}=\frac{{\sum }_1^4{Grazed}_{point(1-4)}}{4}$$

(8)

where

${Grazed}_{point(1-4)}$ are the ADB values for the four grazed sample points.

The difference in ADB was calculated using Equation (9).

$$\mathrm{D}\mathrm{i}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{n}\mathrm{c}\mathrm{e}\_\mathrm{B}\mathrm{i}\mathrm{o}\mathrm{m}\mathrm{a}\mathrm{s}\mathrm{s}=\mathrm{Ungrazed}\_\mathrm{Biomass}-\mathrm{Grazed}\_\mathrm{Biomass}$$

(9)

where

“Ungrazed_Biomass” is the average ADB value of the four ungrazed sample points.

“Grazed_Biomass” is the average ADB value of the four grazed sample points.

Python script served as the primary tool for image processing and analysis, enabling the creation of the time series plots of predicted ADB.

6. Results

6.1. Estimation Model

The biomass estimation model, developed using simple linear regression, correlates ADB from desert, dry, and mountain grasslands with PlanetScope NDVI. Equation (10) represents the predictive line, yielding an R-squared value of 0.6228 and an RMSE of 35.281 g/m² (Figure 12).

y = 309.72x + 8.5032

(10)

where y is the predicted biomass and x is the corrected NDVI of PlanetScope.

Separate correlations between ADB and PlanetScope-corrected NDVI were conducted for desert, dry, and mountain areas, and the regression lines are illustrated in Figure 13. The analysis of correlation gives a correlation coefficient of determination of 0.65 for desert grassland, 0.82 for dry grassland, and 0.80 for mountain grassland.

The outcome from the simple linear regression model suggests that separate modeling for grasslands yields better results than using a single common model.

6.2. Time Series Predicted ADB Mapping Analysis

Utilizing the estimation model (Equation (10)) and the PlanetScope-corrected surface reflectance, biomass maps were created for Mongolia’s most minor administrative unit areas known as ‘Bags’ from 2020 to 2023. These ‘Bag’ areas also include the study sites of the research. The desert grassland area ‘Olon-Ovoo Bag’ in Dalanjargalan soum, Dornogovi province has a total area of 1610 km² (depicted in Figure 14a), the dry grassland area ‘Lkhumbe Bag’ in Tumentsogt soum, Sukhbaatar province covers an area of 1189 km² (shown in Figure 14b), and the mountain grassland area ‘Jargalant Bag’ in Mungunmorit soum, Tuv province encompasses a total area of 1241 km² (as displayed in Figure 14c). All images were selected between August 30 and September 2 and were consistently mapped, categorizing color changes in intervals of 10 g/m² across a biomass range from 0 g/m² to 220 g/m² (Figure 14).

This uniform mapping approach enables meaningful comparative analysis. We also created the histograms of predicted ADB for each map (Figure 15). Figure 14 displays the predicted biomass map images and Figure 15 illustrates the histogram. In the desert grassland, the average recorded ADB was 44 g/m², the maximum mean ADB recorded was 62 g/m² in 2020, while the minimum mean ADB was 33 g/m² in 2022 (Figure 14a and Figure 15a). For the dry grassland, the average recorded biomass was 105.25 g/m². Its maximum mean ADB reached 111 g/m² in 2021, while the minimum mean ADB was 91 g/m² in 2022 (Figure 14b and Figure 15b). In the mountain grassland, the average recorded ADB stood at 133.5 g/m². The maximum mean ADB recorded was 143 g/m² in 2023, while the minimum mean ADB was 120 g/m² in 2021 (Figure 14c and Figure 15c).

Table 3. Predicted ADB for grassland areas.

Value	Type of Grassland	Year				Average (g/m2)
		2020	2021	2022	2023
Mean ADB (g/m2)	Desert grassland	62 RMSE: 15	41 RMSE: 11	33 RMSE: 10	40 RMSE: 11	44 RMSE: 11.75
	Dry grassland	110 RMSE: 14	111 RMSE: 21	91 RMSE: 18	109 RMSE: 18	105.25 RMSE: 17.75
	Mountain grassland	131 RMSE: 24	120 RMSE: 27	140 RMSE: 20	143 RMSE: 20	133.5 RMSE: 22.75

provides the predicted quantitative results for ADB.

The spatial map clearly shows the distribution of biomass and enables us to compare the biomass across different years spatially. For instance, desert grasslands exhibit lower vegetation density in rocky terrain at higher altitudes, whereas areas at lower altitudes with damp conditions support more vegetation (Figure 14a). The histograms provide a comparative quantitative assessment. Notably, 2020 showed significant vegetation growth, while 2022 presented relatively limited vegetation (Figure 15a). In the dry grassland, variations in vegetation growth are evident on either side of the highway, as depicted in Figure 14b. In the mountain grassland, vegetation appears denser compared to the dry and desert areas. Additionally, higher altitude areas exhibit greater vegetation density, as shown in Figure 14c.

6.3. Time Series Grazed and Ungrazed Predicted ADB Dynamic Analysis

Illustrated plot graphs of grazed ADB, ungrazed ADB, and their differences are shown in Figure 16 for desert grassland, Figure 17 for dry grassland, and Figure 18 specifically for mountain grassland.

The grass growth trends align with the results obtained from the mapping analysis by year. For instance, in desert grasslands, there was robust growth in 2020, followed by limited growth in 2022. However, the discrepancies between grazed and ungrazed ADB serve as additional indicators, providing another opportunity to evaluate the grassland.

The trend lines representing these differences are depicted with a red line. In Figure 16, Figure 17 and Figure 18, the trend lines illustrating the differences in ADB values between ungrazed and grazed grassland areas show an increasing trend over time. These trend lines were generated by fitting a linear regression model to the differences in biomass values over time, using the least squares method. This trend potentially suggests the possibility of grassland degradation attributable to overgrazing and human factors.

In the desert grassland, the slope of the difference trend line is 0.12. In the dry grassland area, the slope of the difference trend line is 0.06. Lastly, in the mountain grassland, the slope of the difference trend line is 0.03.

7. Discussion

Biomass estimation model in grasslands: The biomass estimation model developed in this study, utilizing simple linear regression with satellite NDVI data and field-measured ADB data, yielded a coefficient of determination (R²) of 0.62 in the combined three grassland areas. However, when analyzed separately, the correlation coefficients of determination were 0.65 for desert grassland, 0.82 for dry grassland, and 0.80 for mountain grassland. These findings suggest that separate models should be considered for these three distinct grassland types.

Comparing our results with recent studies on biomass estimation in Mongolian grasslands, we acknowledge the superior performance of models utilizing high-resolution satellite imagery such as Sentinel-2 and Landsat 8, which offer pixel resolutions of 10 m and 30 m, respectively [13,14]. These studies demonstrated higher coefficients of determination (R² ≈ 0.75) and lower root-mean-square errors (RMSE ≈ 20 g/m²) by employing advanced methodologies like Partial Least Squares (PLSs) and Random Forest (RF) methods, integrating multiple parameters including the Green Chlorophyll Index, Enhanced Vegetation Index, Simple Ratio, Wide Dynamic Range Vegetation Index, and NDVI. Furthermore, the potential of PlanetScope satellite data for biomass estimation has been demonstrated in diverse environments, such as coastal wetlands in South Carolina, achieving an R² of 0.75 [31].

Our study aimed to assess the applicability of PlanetScope imagery for biomass estimation in Mongolia. We employed a single linear regression model with NDVI as the primary parameter, a commonly used metric for biomass estimation in grassland environments. While our findings are consistent with prior research indicating a correlation coefficient of determination (R² = 0.68) between NDVI and biomass [32], the relatively low R² obtained in our study suggests room for improvement. Additionally, our methodology involved satellite data correction tailored to the study area, leading to improved accuracy, particularly noticeable in the desert grassland (

Table 4. Comparison between data before and after radiometric correction.

	NDVI and Biomass Correlation (R2)				RMSE of Combined Biomass (g/m2)
	In Desert Grassland	In Dry Grassland	In Mountain Grassland	Combined Grassland
Before radiometric correction	0.626	0.827	0.814	0.6227	35.282
After radiometric correction	0.656	0.822	0.8	0.6228	35.281

). This observation may also suggest the distinctive characteristics of the desert region.

In summary, our biomass estimation model highlights the potential of high-resolution satellite imagery in Mongolian grasslands. However, further research is warranted to develop more robust models that consider a broader range of parameters and address the limitations of optical satellite imagery. These limitations include the presence of shadows and clouds, geographical accuracy, and the reliance on a single parameter for correlation. Moreover, ground field measurements necessitate significant time and cost resources, which should be considered in future studies.

Time series mapping and dynamic analyses in grasslands: We conducted time series analyses on grassland areas using predicted biomass, facilitated by high-resolution satellite data. The revised daily observation interval improved our data collection capabilities, although scaling this approach to larger regions like Mongolia requires substantial computational resources.

These analyses span the last four years (2020–2023) and mapping analyses were conducted with biomass maps produced for the same day of each year. As illustrated on the map, the time series maps demonstrate varying grass growth from year to year, significantly influenced by weather conditions and climate. For example, desert grasslands experienced high growth in 2020, decreased in 2022, then rebounded in 2023. Additionally, these maps have demonstrated the feasibility of examining the impact of infrastructure, such as highways, on grasslands in specific areas.

Secondly, based on the high-resolution constellation satellite’s advantage, we conducted grazed and ungrazed vegetation dynamic analysis, utilizing our predicted biomass for ungrazed and grazed areas. This analysis result indicates that the disparities between grazed and ungrazed ADB have been observed to increase over time. This escalation varies across different regions; a rapid increase is noted in the desert area, a moderate increase in the dry area, and a slower rise in the mountainous region. This comparison serves as a potential indicator of the grassland degradation status. In a previous study, the National Report on Grazing Impact Monitoring in Mongolia highlighted that desert grasslands are particularly prone to degradation and are significantly influenced by climate change [11].

For dynamic analyses, only satellite data were used, covering only the growing season and not all seasons. It may be beneficial to include other parameters such as weather data, livestock numbers, and population demographics to better explain these changes. Additionally, the analyses were conducted in only one area within each of the three grassland regions, which may not fully represent each area.

In Figure 16, Figure 17 and Figure 18, especially in Figure 16, some points show sharp changes, possibly influenced by wet soil (rain) and limitations of optical satellite data. However, these changes are evident in both ungrazed and grazed areas so, when focusing on their differences, it is acceptable.

The results of the time series dynamic analysis suggest that decision-makers should prioritize attention to desert and dry areas over mountainous regions. However, it is important to acknowledge that degradation occurs in all areas.

8. Conclusions

This study successfully utilized PlanetScope high-resolution satellite data for estimating biomass in Mongolian grasslands using a simple linear regression model, along with grassland vegetation time series mapping and dynamic analyses. The results highlight the following key findings: 1. The analysis demonstrated the capability of PlanetScope data in estimating biomass, with a correlation coefficient of determination of 0.62 when combined across all grasslands. When analyzed separately, the correlation coefficients were 0.65 for desert grassland, 0.82 for dry grassland, and 0.80 for mountain grassland. These results suggest that separate models for specific grassland types may provide better results. 2. The assessment of PlanetScope data accuracy in the Mongolian area revealed geographic coordinate measurements ranging from 2 to 6 m. Radiometric correction achieved an R² of 0.97 for the red and NIR bands. 3. The mapping and grazed and ungrazed dynamic analyses conducted using high-resolution satellite data unveiled operational insights into grassland dynamics. Spatial differences in vegetation growth were vividly illustrated, while dynamic analysis indicated ongoing degradation processes within Mongolian grasslands, with rapid degradation observed in desert grassland, moderate degradation in dry grassland, and low degradation in mountain grassland.

In the future, we aim to enhance our estimation model by integrating multiple parameters with PlanetScope data and employing machine learning algorithms in our upcoming projects. Following this, our objective is to develop a monitoring platform that utilizes multi-satellite validation with our model. Accurate estimation of grassland parameters is crucial for evaluating their condition and facilitating ongoing monitoring processes.

This study contributes to the growing body of research on remote sensing applications in grassland ecology and underscores the importance of leveraging high-resolution satellite data for improved environmental assessment and management practices.