Adapting the High-Resolution PlanetScope Biomass Model to Low-Resolution VIIRS Imagery Using Spectral Harmonization: A Case of Grassland Monitoring in Mongolia

Abstract

Monitoring grassland biomass accurately and frequently is critical for ecological management, climate change assessment, and sustainable resource use. However, the use of single-satellite data faces challenges due to trade-offs between spatial resolution and temporal frequency, especially for large areas. High-resolution imagery, such as PlanetScope, provides detailed spatial data but presents significant challenges in data management and processing over large regions. Conversely, low-resolution sensors such as JPSS-VIIRS offer daily global coverage with low memory data but lack the spatial detail required for precise biomass estimation, making it difficult to retrieve or validate model parameters due to the mismatch with small ground reference data polygons. To overcome these limitations, this study introduces a robust methodology for accurate frequent biomass estimation based on JPSS-VIIRS data through spectral harmonization, adapting a high-resolution biomass estimation model originally developed from PlanetScope imagery. The core innovation is an optimized Spectral Band Adjustment Factor (SBAF) approach tailored specifically to grassland spectral characteristics. This method significantly enhances spectral alignment, reducing red-band reflectance discrepancies from 6.2% to 4.8% in grassy areas and from 6.9% to 4.0% in bare areas. NDVI discrepancies also improved substantially. Applied across Mongolia, the harmonized VIIRS data estimated a five-year average biomass of 71.4 g/m², clearly reflecting environmental variability. Specifically, the P375 dataset showed average biomass estimates of 54.8 g/m² for desert grasslands (10.5% higher than PlanetScope), 122.6 g/m² for dry grasslands (9.6% higher), and 134 g/m² for mountain grasslands (1.9% lower). The uncertainty analysis showed strong overall agreement with PlanetScope-derived biomass, with an RMSE of 11.6 g/m², a mean percentage difference of 10.74%, and an R² of 0.92. While mountain grasslands exhibited the lowest RMSE, a relatively lower R² indicated limited variability. Higher uncertainty in desert and dry grasslands highlighted the impact of ecological heterogeneity on biomass estimation accuracy. These detailed comparisons demonstrate the effectiveness and accuracy of the proposed methodology in bridging spatial and temporal gaps, providing a valuable tool for large-scale weekly grassland biomass monitoring with applicability beyond the Mongolian context.

1. Introduction

Advances in Earth observation sensors have significantly increased the capacity for more frequent and accurate terrestrial monitoring, enabling a broader range of practical applications [1,2]. One critical application is the spatial monitoring of grassland growth, which often relies on multiple sensor images or diverse datasets to cover a large geographic area [3,4]. However, achieving accurate parameter estimation and near-real-time observations remains challenging when integrating data from different sensors and acquisition dates. Therefore, daily imagery with sufficient coverage and robust modeling accuracy from a single sensor is essential for effective grassland growth monitoring at regional and global scales.

The Moderate Resolution Imaging Spectroradiometer (MODIS) has historically been a key tool for grassland monitoring at these scales [4,5,6]. While MODIS offers excellent temporal resolution, its coarse spatial resolution limits the precision and reliability of research outcomes [3,7,8]. This drawback is particularly pronounced because the ground sampling data used for modeling are often much smaller in scale (e.g., 1 m × 1 m) than MODIS’s pixel size [9,10]. Additionally, as the aging MODIS satellites, operational since 1999, approach the end of their lifespans, the Visible Infrared Imaging Radiometer Suite (VIIRS) has assumed similar roles in Earth observation [11].

Recent advancements in satellite technology have made higher-quality imagery increasingly accessible for grassland monitoring. Satellites such as Landsat and Sentinel-2 provide higher spatial resolution than MODIS, improving the accuracy of biomass and other parameter estimation models [3,8,12,13,14,15]. However, these satellites face limitations, including individual revisit cycles of 16 days for Landsat (reduced to ~8 days with Landsat-8 and Landsat-9 combined) and 10 days for Sentinel-2 (reduced to ~5 days with Sentinel-2A and Sentinel-2B combined), alongside frequent cloud cover that disrupts data collection [16]. Over the past decade, small-constellation satellites such as PlanetScope have revolutionized this field by offering daily high-resolution imagery with a spatial resolution of 3 m [17]. While high-resolution data significantly enhances modeling precision, it poses challenges related to the substantial storage and processing requirements for regional and global areas. For grassland biomass estimation, various methods have been successfully applied, including multiple regression analysis [18], artificial neural networks (ANN) [19], and K-nearest neighbor [20]. However, in this study, a simple linear regression model was selected as the basis for model adaptation due to its interpretability and suitability for large-scale applications.

A promising method for increasing observation frequency involves using virtual constellations (VC), which integrate and harmonize data from multiple satellites to improve revisit times [21]. However, despite having similar bands, satellite sensors often differ in spectral ranges and RSR, leading to discrepancies in measured values when observing the same target under identical conditions [22]. This necessitates effective spectral harmonization. Various techniques have been developed to correct spectral differences, including linear models for Landsat and the Harmonized Landsat Sentinel-2 (HLS) product [23], the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), which blends Landsat and MODIS reflectance data to produce daily 30 m resolution imagery [24], and the SBAF technique, which uses hyperspectral data to characterize the relationship between sensors’ RSR [25].

The evaluation of spectral harmonization techniques has shown that multi-band input methods yield better results than single-band approaches, with surface characteristics playing a significant role [26]. The SBAF technique, which is based on a single band, has been suggested for globally mixed areas in previous studies [25,26,27,28]. It is derived from actual hyperspectral data and sensor RSRs. For instance, the SBAF estimation online tool for multiple sensors uses hyperspectral SCIAMACHY data [27], while Earth Observation-1 (EO-1) Hyperion datasets have been used for SBAF estimation [28]. Recent developments have also explored numeric and machine-learning-based approaches for harmonization, such as random forest regression, support vector machines, and deep learning frameworks [25,26,27,28]. These methods show promise but have not been specifically optimized for grassland applications.

This research targets grassland areas where Red-NIR reflectance is highly sensitive. For example, the VIIRS red band covers a spectral range of 600 to 680 nm, while the PlanetScope SuperDove red band spans 650 to 680 nm. Grassland reflectance behavior within this range varies: in greener (grassy) areas, reflectance increases between 650 and 680 nm, while, in less green (bare) areas, reflectance decreases. Addressing these reflectance characteristics, the optimized SBAF technique provides an effective and straightforward approach for harmonizing between sensors.

This study introduces a methodology for large-scale grassland biomass monitoring by adapting a high-resolution biomass estimation model, originally developed from PlanetScope imagery, to low-resolution JPSS-VIIRS data through spectral harmonization. The approach utilizes an optimized SBAF technique, tailored to grassland’s spectral characteristics, to enable accurate and frequent monitoring over expansive regions. Unlike traditional harmonization methods such as linear regression or machine learning approaches, which often lack ecosystem-specific tuning, our optimized SBAF leverages multi-source hyperspectral data to define NDVI-based selection thresholds, reducing red-band reflectance discrepancies from 6.2% to 4.8% in grassy areas and from 6.9% to 4.0% in bare areas. Applied across Mongolia from 2020 to 2024, the harmonized VIIRS (P375) data estimated a five-year average biomass of 71.4 g/m², with temporal variation reflecting environmental conditions. Specific estimates included 54.8 g/m² for desert grasslands, 122.6 g/m² for dry grasslands, and 134 g/m² for mountain grasslands, with an overall RMSE of 11.6 g/m² and an R² of 0.92 against PlanetScope-derived biomass. A comparative analysis of optimized and non-optimized SBAF techniques across Mongolia’s diverse grasslands confirmed the superior performance of the optimized approach, offering a scalable solution for weekly biomass monitoring beyond this region.

2. Methodology

This section outlines the workflow of the study, which introduces a novel approach for large-scale grassland biomass monitoring using VIIRS imagery adapted through a high-resolution PlanetScope-based biomass model [29]. The methodology integrates all necessary steps to ensure temporal consistency and broad spatial coverage, ultimately enabling weekly biomass estimation across Mongolia. The overall workflow consists of three main components: data collection and pre-processing, model adaptation, and processing and analysis, as illustrated in Figure 1.

The processing chain consists of four steps:

• Data collection and Preprocessing: This study utilized data from JPSS-2 VIIRS, PlanetScope SuperDove, and EO-1 Hyperion satellites, along with Relative Spectral Response (RSR) profiles for both VIIRS and PlanetScope sensors. Additionally, ground-based field reflectance measurements were incorporated to support the optimization of SBAF.
  • VIIRS Preprocessing: the VIIRS Sensor Data Record (SDR L1B) was converted from TOA to BOA reflectance, georeferenced to a projected coordinate system, and composited into a 7-day max NDVI cloud-free mosaic suitable for grassland monitoring.
  • PlanetScope data: PlanetScope surface reflectance data were used directly to evaluate spectral harmonization.
  • Hyperion Data: Hyperspectral surface reflectance data from EO-1 Hyperion were combined with VIIRS and PlanetScope RSR profiles to simulate sensor-specific reflectance values.
  • Field Measurements SR: Ground-based reflectance measurements, including in situ field campaigns and RadCalNet-BSCN site observations, were integrated with Hyperion data to improve SBAF optimization.
• Model Adaptation: This model adaptation process enables the transfer of high-resolution biomass estimation models developed from PlanetScope imagery to daily VIIRS data through optimized spectral harmonization.
  • Development of SBAF: SBAFs were calculated using Hyperion hyperspectral data and the RSR profiles of PlanetScope and VIIRS, capturing spectral differences between the two sensors.
  • Optimization of SBAF for Grassland: SBAFs were optimized by classifying land surfaces into grassy and bare areas using NDVI thresholds, improving accuracy by addressing surface-specific spectral variability.
  • Spectral Harmonization: Optimized SBAFs were applied to VIIRS reflectance data to harmonize it with PlanetScope, enabling the adapted biomass model to be used on VIIRS imagery.
• Processing and Analysis: This step applies the harmonized data in the biomass estimation workflow, enabling regional-scale grassland biomass estimations and evaluations of the harmonization performance.
  • Biomass Estimation with Harmonized VIIRS: The harmonized VIIRS reflectance (P375) was used as the input to the PlanetScope-based biomass model to produce daily biomass estimates.
  • Evaluation of SBAF Optimization: This evaluation involved comparing PlanetScope surface reflectance with both the original and the harmonized VIIRS reflectance (after applying SBAF). By assessing differences before and after harmonization, the effectiveness of SBAF optimization was quantified across grassland areas.
• Output (Results):
  • Biomass Estimation and Mapping: Weekly biomass and NDVI maps were generated from harmonized VIIRS data (P375), providing spatially continuous and temporally consistent representations of grassland biomass dynamics across Mongolia over multiple years.
  • Spectral Harmonization for Grassland: Spectrally aligned VIIRS surface reflectance data were produced using optimized SBAFs derived from Hyperion hyperspectral data and ground-based measurements, ensuring compatibility with high-resolution PlanetScope-based biomass models.

3. Study Area

The study focuses on Mongolia, part of the broader Eurasian semi-arid grassland region, which is classified as temperate grasslands [30]. Mongolia’s grasslands dominate the country’s landscape, with approximately 110.3 million hectares serving as pastureland, making it a critical area for wild grazing. These grasslands represent a significant portion of Mongolia’s 156.3 million hectares of land and form a vital link in the ecological and climatic dynamics of the Eurasian grasslands. Together with the 1.7 million hectares allocated for agriculture, including grasslands and crop areas, they account for roughly 71.8% of Mongolia’s total area [31,32].

Figure 2 illustrates a true-color composite of Mongolian VIIRS imagery. The triangular markers represent surface reflectance (SR) field measurement (FM) points, with a green triangle indicating the mountain grassland (MG) measurement point, a blue triangle representing the dry grassland (DG) measurement point, an orange triangle representing the desert grassland (DesG) measurement point, a red triangle representing the sand measurement point, and green circular markers indicating the locations of spectral harmonization evaluation points.

4. Data Collection and Preprocessing

This study utilized both satellite imagery and ground-based surface reflectance (SR) data. The satellite datasets included low-resolution JPSS-VIIRS, high-resolution PlanetScope SuperDove SR, and hyperspectral EO-1 Hyperion SR. These datasets were used for biomass model adaptation, spectral harmonization between VIIRS and SuperDove in grasslands, SBAF optimization, and the mapping of NDVI and aboveground biomass.

Ground-based SR measurements were used to characterize grassland reflectance properties in the red and NIR bands. These field data, combined with Hyperion SR, supported the optimization of the SBAF techniques.

4.1. Joint Polar-Orbiting Satellite System (JPSS)—Visible Infrared Imaging Radiometer Suite (VIIRS) Imagery and Preprocessing (Geo-Rectified Surface Reflectance)

The Joint Polar Satellite System (JPSS), developed by NOAA and NASA, is a next-generation polar-orbiting environmental satellite system. Its Visible Infrared Imaging Radiometer Suite (VIIRS), onboard the JPSS—Suomi NPP, JPSS1—NOAA-20, and JPSS2—NOAA-21 satellites, launched between 2011 and 2022, plays a vital role in global environmental monitoring and weather forecasting. VIIRS builds on the capabilities of AVHRR, MODIS, and Sea-WiFS by observing clouds and Earth surface variables, while other instruments onboard focus on atmospheric and radiation measurements. With a field of view of 112.56°, VIIRS offers a swath width of 3060 km at an altitude of 829 km, providing full daily coverage of the Earth, both day and night. The instrument contains 22 spectral bands: 16 moderate-resolution bands (M-bands) with a 750 m spatial resolution at the nadir, five imaging bands (I-bands) with a 375 m spatial resolution, and a panchromatic day/night band (DNB) with a 750 m resolution [11].

For this study, 140 granules of VIIRS Sensor Data Record (SDR) data, spanning 35 days from 1–7 September over the past five years (2020–2024), were sourced from NOAA’s CLASS Archive to map the Mongolian area. Approximately four granules per day were required to ensure complete coverage of Mongolia. The SDR data, formatted in Hierarchical Data Format version 5 (HDF5), included top-of-atmosphere (TOA) reflectance, radiance, and brightness [11].

The VIIRS SDR data required pre-processing, including converting TOA to bottom-of-atmosphere (BOA) reflectance, georeferencing, and creating seven-day cloud-free composites for the purposes of this study.

4.1.1. VIIRS TOA Data Conversation to BOA Reflectance and Georeferencing

The processing workflow involved converting VIIRS SDR Level-1 TOA reflectance to BOA reflectance and georeferencing. This general process diagram is illustrated in Figure 3, with the software workflow detailed in Figure 4. The Community Satellite Processing Package (CSPP), developed by the Space Science and Engineering Center (SSEC) at the University of Wisconsin–Madison, was employed for this purpose [33].

The conversion to BOA reflectance was conducted in two steps:

• The process begins by generating the required auxiliary data files, including cloud masks, cloud height, aerosol properties, and weather prediction data. This is achieved using the CSPP VIIRS ASCI V1.2 software, which processes VIIRS SDR and dynamic ancillary data [33].
• In the second step, the conversion to BOA reflectance is performed using the CSPP VIIRS Land Surface Reflectance V1.1 software. This process involves applying the Lambertian approximation for atmospheric correction, adjacency adjustments to reduce glare from surrounding pixels, and bidirectional reflectance distribution function (BRDF) coupling adjustments. It leverages the auxiliary data generated in the first step and the SDR data [33,34].

The georeferencing process involved remapping, gridding, and projection. Nearest-neighbor remapping was applied with a spatial resolution of 375 m, focusing on the Mongolian region. The data were projected to the WGS84 coordinate reference system and saved as GeoTIFF images. These operations were performed using Polar2Grid command line V3.1 software [33].

4.1.2. Maximum NDVI Cloud-Free Seven-Day Composite

To facilitate accurate, precise mapping and analysis, a maximum NDVI composite was generated for each pixel across the Mongolian region over 35 days spanning September 1–7 for the past five years (2020–2024). Daily NDVIs were calculated, and the maximum value within each seven-day window was selected for each pixel. This method inherently excluded cloud-contaminated pixels, as clouds typically exhibit lower NDVI values [35]. A water mask was also applied to delineate water bodies, ensuring that the final composite accurately supported grassland mapping, the evaluation of spectral harmonization, and overall analysis within the study area.

4.2. PlanetScope—SuperDove Imagery

The PlanetScope constellation, which includes Dove, Dove-R, and SuperDove sensors, provides high spatial resolution imagery at 3.125 m with 4 or 8 bands in the visible-to-NIR range, offering daily commercial satellite data [17].

For this study, 30 scenes captured by the third-generation SuperDove sensor between September 1 and 7 over the past two years (2023–2024) were selected for spectral harmonization evaluation with VIIRS imagery. These scenes were acquired exclusively through the Planet Explorer platform (https://www.planet.com/explorer/, accessed on 21 October 2024). The selection criteria prioritized clear-sky conditions to minimize cloud contamination, ensuring accurate SR data for evaluation and analysis.

4.3. The Earth Observation-1 (EO-1)—Hyperion Hyperspectral Surface Reflectance

The Hyperion sensor aboard the Earth Observing-1 (EO-1) satellite, operational from 2000 to 2017, provides a unique archive for spectral libraries under real-world conditions. Hyperion captures data across 220 spectral bands (357 nm to 2.576 µm) with a spatial resolution of 30 m and a 7.7 km swath [36].

For this study, the Hyperion surface reflectance dataset, provided by the European Commission’s Joint Research Centre, includes georeferenced and atmospherically corrected surface reflectance measurements from 10,000 points across various global surface types [28,37]. This dataset serves as a valuable resource for surface reflectance analysis and was used for simulated surface reflectance to estimate the SBAF in spectral harmonization.

4.4. Ground Field Surface Reflectance Data

Ground field data (Figure 5) were collected to analyze grassland reflectance characteristics and identify the threshold NDVI at the transition in reflectance. This transition refers to the shift from VIIRS Red band reflectance being greater than or equal to PlanetScope Red band reflectance (VIIRS_Red ≥ PlanetScope_Red), as illustrated in the “SR Hyperion Grassy” graph in Figure 5, to VIIRS Red band reflectance becoming less than PlanetScope Red band reflectance (VIIRS_Red < PlanetScope_Red), as shown in the “SR Hyperion Bare” graph in Figure 5. This change underscores the sensitivity of the red band reflectance in grasslands. Additionally, the threshold NDVI for sand was identified.

Field measurements were conducted from 26 July to 1 August 2023, across three grassland types in Mongolia: mountain, dry, and desert grasslands [29]. An ASD FieldSpec HandHeld 2 Spectroradiometer (325–1025 nm, 25° FOV) was used. Measurements were taken at a height of 1.2 m, covering a 0.23 m² ground area. A total of 384 spectral measurements were collected from 24 plots. Sand reflectance data were sourced from RadCalNet’s Baotou Comprehensive Calibration and Validation Site (BSCN) [38,39].

All field measurements were performed on proximate dates at the exact times and locations marked by the triangles in Figure 2. The resulting surface reflectance graphs, including the observed transitions and sand reflectance, are presented in Figure 5.

5. Model Adaptation

To support consistent and timely large-scale grassland biomass estimation, this section describes the adaptation of a high-resolution biomass estimation model, originally developed using PlanetScope imagery, for use with daily VIIRS data via optimized spectral harmonization.

5.1. Development of SBAF

The development of SBAF follows the process outlined in Figure 6, which involves spectral harmonizing VIIRS SR with PlanetScope SR to minimize differences caused by variations in sensor RSR.

To simulate surface reflectance (${\overline{\rho }}_{\lambda }$) for a satellite sensor’s broader spectral band, hyperspectral surface reflectance (${\rho }_{\lambda }$) was weighted according to the sensor’s RSR and integrated over the bandpass (Equation (1)) [25,26].

$${\overline{\rho }}_{\lambda }={\displaystyle \frac{\int {\rho }_{\lambda }\times RS{R}_{\lambda }\times \mathrm{\Delta }\lambda }{\int RS{R}_{\lambda }\times \mathrm{\Delta }\lambda }}$$

(1)

Hyperion’s spectral bands encompass narrow wavelength (10 nm) intervals, with the RSR exhibiting variation within these intervals following a non-linear function (Figure 5). The RSR is modeled (Equation (2)) using a Gaussian function, characterized by the average wavelength (μ) and the full width at half maximum (Equation (3)) [26,28]:

$$RS{R}_{\lambda }={\displaystyle \frac{1}{\sigma \sqrt{2\pi }}}{e}^{{\displaystyle \frac{{-(\lambda -\mu )}^2}{2{\sigma }^2}}}$$

(2)

where σ is defined as:

$$\sigma ={\displaystyle \frac{FWHM}{2\sqrt{2\log \left(2\right)}}}$$

(3)

The integration of Hyperion’s hyperspectral reflectance profile was calculated using the Hyperion reflectance (${\rho }_i$), the sensor’s RSR ($RS{R}_b$), and the relative spectral response of each Hyperion spectral band ($RS{R}_H$) (Equation (4)):

$${\overline{\rho }}_b={\displaystyle \frac{{\sum }_i({\rho }_i\int (RS{R}_{H,j}\left(\lambda \right)\times RS{R}_{b,j}\left(\lambda \right)\mathrm{\Delta }\lambda ))}{{{\displaystyle \sum }}_i(\int \left(RS{R}_{H,j}\left(\lambda \right)\times RS{R}_{b,j}\left(\lambda \right)\mathrm{\Delta }\lambda \right))}}$$

(4)

where ${\overline{\rho }}_b$ is the simulated reflectance of spectral band b with the corresponding RSR ($RS{R}_b$), i—Hyperion spectral band, and j—RSR band range of Hyperion.

Once the band’s simulated SR is identified, the SBAF is applied according to Equation (5):

$$\mathrm{SBAF}={\displaystyle \frac{{\overline{\rho }}_{\lambda (b\_\mathrm{tar})}}{{\overline{\rho }}_{\lambda (\mathrm{b}\_\mathrm{ref})}}}$$

(5)

where ${\overline{\rho }}_{\lambda (b\_\mathrm{tar})}$ is the simulated SR of the target sensor’s band (VIIRS), and ${\overline{\rho }}_{\lambda (b\_ref)}$ is the simulated SR of the reference sensor’s band (PlanetScope-SuperDove).

5.2. Optimization of SBAF for Grassland

This study focused on the Red and NIR bands, which are highly sensitive to vegetation reflectance. Due to there being broader spectral coverage in the VIIRS Red band (600–680 nm) compared to PlanetScope (650–680 nm), as illustrated in

Table 1. Super-Dove and VIIRS sensor specification.

Instrument	Red Band Range	NIR Band Range	Spatial Resolution	Revisit Time
PlanetScope SuperDove	650–680 nm	845–885 nm	3.125 m	Daily
JPSS-2 VIIRS	600–680 nm	845–885 nm	375 m	Daily

, greater variability was observed in reflectance, particularly in the red band. As shown in Figure 5, grassy areas typically exhibit decreasing red reflectance toward longer wavelengths, while bare areas show the opposite trend. This spectral behavior informed the optimization of SBAFs.

To address surface-specific variability, SBAFs were optimized by classifying the reflectance conditions into three categories: grassy, bare, and sandy surfaces.

• Grassy grassland: NDVI ≥ 0.3 (VIIRS red reflectance ≥ PlanetScope red reflectance)
• Bare grassland: 0.12 < NDVI < 0.3 (VIIRS red reflectance < PlanetScope red reflectance)
• Sand: NDVI ≤ 0.12 (no grassland)

The threshold NDVI values were derived by averaging the NDVI, calculated as follows:

• Hyperion reflectance-based classification (grassy mean NDVI_grassy ≈ 0.568 (3506 points), bare mean NDVI_bare ≈ 0.237 (6494 points))
• Field-based observations (desert grassland NDVI_grassland ≈ 0.198)
• RadCalNet sandy site (NDVI_sand ≈ 0.12)

The classification threshold NDV_hyperion is 0.402 from the Hyperion SR:

$$NDV{I}_{hyperion}={\displaystyle \frac{NDV{I}_{grassy}+NDV{I}_{bare}}{2}}=0.402$$

(6)

Finally, based on these NDVIs, the classification threshold NDVI_class = 0.3 was calculated using Equation (7).

$$NDV{I}_{class}={\displaystyle \frac{NDV{I}_{Hyperion}+NDV{I}_{grassland}}{2}}=0.3$$

(7)

The resulting classification threshold of NDVI = 0.3 was used to assign optimized SBAFs to each pixel when harmonizing the VIIRS data. This improved spectral alignment and model accuracy across grassland types.

5.3. Spectral Harmonization

The spectral harmonization of VIIRS imagery to PlanetScope imagery was conducted using the following steps, as illustrated in Figure 7. First, NDVI was estimated. Second, SBAF optimization was applied. Finally, SBAFs were applied for spectral harmonization. This process was implemented on a pixel-by-pixel basis.

6. Processing and Analysis

6.1. Biomass Estimation with Harmonized VIIRS Data and Mapping

Biomass estimation was conducted using a high-resolution PlanetScope imagery-based model (Equation (8) [29]) after radiometric correction [29]. Subsequently, harmonized VIIRS biomass estimation was applied to three major Mongolian grassland types—desert steppe, dry steppe, and mountain steppe—and extended to the entire territory of Mongolia. PlanetScope-derived biomass served as a ground-truth reference for comparison with biomass estimates derived from harmonized VIIRS data:

Biomass = 309.72 × NDVI_viirs + 8.5032

(8)

To systematically evaluate the uncertainty of the VIIRS-based biomass estimates, a detailed comparison with PlanetScope-derived biomass data was conducted for three principal grassland types—desert, dry, and mountain—in selected small administrative units (Bags) across Mongolia (2020–2024) [29]. Specifically:

• Desert Grassland: Olon-Ovoo Bag, Dalanjargalan soum, Dornogovi province (1610 km²)
• Dry Grassland: Lkhumbe Bag, Tumentsogt soum, Sukhbaatar province (1189 km²)
• Mountain Grassland: Jargalant Bag, Mungunmorit soum, Tuv province (1241 km²)

Uncertainty was evaluated using statistical metrics, including the root mean square error (RMSE), mean percentage difference, and coefficient of determination (R²), to quantify the reliability of VIIRS-based biomass estimations under varying ecological conditions.

The biomass mapping process for the entire Mongolian region utilized seven-day maximum NDVI cloud-free mosaic images collected from 2020 to 2024, following methodologies outlined in previous research [29]. The mapping workflow consisted of four main steps:

• Harmonization and corrections: Original VIIRS SR data were harmonized, and sensor-specific adjustments were made to address radiometric discrepancies for grassland [29].
• NDVI estimation: NDVI was calculated using corrected red and NIR bands.
• Biomass estimation: biomass was then derived from the corrected NDVI, as shown in Equation (8).
• Map generation: final biomass and NDVI maps were produced. A water mask was applied to exclude water bodies and ensure a proper focus on grassland areas.

6.2. Evaluation of SBAF Optimization

The SBAFs were derived separately for grassy and bare areas in the red and NIR bands, drawing on the coefficients. Two spectral harmonization techniques were then tested: an optimized SBAF and a non-optimized SBAF.

To evaluate these techniques, near-nadir VIIRS SR data were collected for 593 selected grassland points in Mongolia, focusing on a cloud-free period during the first week of September (2023–2024). PlanetScope data, featuring a 3.125 m spatial resolution, were used as a reference. Because the VIIRS pixel size is 375 m, each corresponding 375 m × 375 m area of PlanetScope imagery was averaged to approximate one VIIRS pixel. This averaging process relied on determining the center coordinate of each VIIRS pixel, which then served as the center point for extracting and averaging the 15,625 PlanetScope pixels (i.e., 125 × 125 pixels) that fit within the same spatial footprint.

By comparing the aggregated PlanetScope data to both the original and harmonized VIIRS SR (i.e., before and after SBAF optimization), reflectance differences in red, NIR, and NDVI were calculated as percentages. This allowed for a direct assessment of how effectively the SBAF optimization technique minimized discrepancies between VIIRS and PlanetScope measurements across varying surface conditions (grassy vs. bare areas).

7. Results

7.1. Biomass Estimation and Mapping

The uncertainty evaluation of harmonized VIIRS biomass estimates was conducted through detailed comparisons with PlanetScope-derived biomass values (ground truth) for the three primary grassland types—desert, dry, and mountain—in selected administrative units across Mongolia from similar days of 2020 to 2024. The comparison quantified the accuracy using root mean square error (RMSE), mean percentage difference, and coefficient of determination (R²). The results showed strong overall agreement between harmonized VIIRS-based estimates and PlanetScope-derived values, with all grasslands combined yielding an RMSE of 11.6 g/m², a mean percentage difference of 10.74%, and an R² of 0.92. Although mountain grasslands had the lowest RMSE, the relatively low R² suggests limited variability in the observed biomass values. Desert and dry grasslands exhibited higher uncertainty, reflecting the influence of ecological variability on biomass estimation performance using harmonized VIIRS data.

Biomass and NDVI maps generated from the harmonized VIIRS data visually depict spatial and temporal biomass dynamics across Mongolia. The maps reveal clear spatial variability in biomass, influenced by ecological and climatic conditions. Biomass trends across Mongolia showed substantial temporal variation over the study period. The highest average biomass was observed in 2024 (78 g/m²), likely due to favorable climatic conditions, while 2022 recorded the lowest average biomass (60 g/m²), reflecting potential environmental stress. The five-year mean biomass was calculated as 71.4 g/m². These findings provide valuable insights for grassland management and monitoring. Figure 8 presents the final biomass and NDVI maps.

7.2. Spectral Harmonization Performance

SBAFs were derived separately for grassy and bare areas in the red and NIR bands, as illustrated in

Table 2. SBAFs for JPSS-VIIRS to PlanetScope-SuperDove.

	Bands	SBAF
Point Type		Red	NIR
Optimized SBAF	Grassy (440 point)	0.950318	1.001061037
	Bare (153 point)	1.037896861	1.000992597
Non-optimized SBAF (all 593 points)		1.008146241	1.001015846

. Two spectral harmonization approaches—optimized SBAF and non-optimized SBAF—were compared to evaluate their effectiveness.

The evaluation utilized real satellite data, comparing VIIRS surface reflectance (SR) before and after SBAF application against high-resolution PlanetScope SR data. The analysis included 593 grassland points across Mongolia using cloud-free imagery acquired during the first week of September for 2023 and 2024, selecting near-nadir VIIRS imagery to ensure optimal quality.

Spatial alignment was necessary due to resolution differences (375 m for VIIRS vs. 3.125 m for PlanetScope). VIIRS pixels were spatially matched by averaging the 15,625 corresponding PlanetScope pixels.

Before spectral harmonization, substantial reflectance discrepancies were observed, particularly in the Red and NDVI bands, across both grassy and bare areas. Following the application of classified SBAFs, these discrepancies notably decreased, especially in grassy areas. For instance, NDVI differences in grassy areas improved from 5.5% before harmonization to 3.1% afterward (Figure 9 and

Table 3. Evaluation of harmonization.

	Bands	Percentage Difference Before and After SBAF Harmonization
		Red		NIR		NDVI
Point Type		Before	After	Before	After	Before	After
Optimized SBAF	Grassy (440 point)	6.2	4.8	3.9	3.8	5.5	3.1
	Bare (153 point)	6.9	4	4.8	4.7	6.1	4.9
Non-optimized SBAF (all 593 points)		6.4	6.6	4.1	4.1	5.6	5.9

). Improvements were less pronounced in bare areas, suggesting that classified SBAFs more effectively reduced discrepancies in vegetated surfaces.

These findings highlight the value of optimized SBAF in enhancing spectral compatibility between VIIRS and PlanetScope imagery, though further refinement might be necessary for bare surfaces.

8. Discussion

This study demonstrates that the high-resolution PlanetScope biomass estimation model [29] can be adapted to VIIRS imagery through spectral harmonizing, offering a temporal and accurate solution for grassland monitoring. The spectral harmonization process, implemented using the SBAF technique, addresses discrepancies caused by variations in sensor RSR and is derived from hyperspectral Hyperion SR. This section explores the implications of the findings, identifies limitations, and proposes future research directions.

Advancements in Grassland Biomass and NDVI Mapping: The adaptation of the PlanetScope biomass estimation model provides a feasible and scalable solution for improving estimation accuracy in low-resolution imagery, enabling more effective broad-scale grassland monitoring. This study emphasized the adaptation process and spectral harmonization rather than directly evaluating the biomass estimation model itself. Previous research has demonstrated that high-resolution remote sensing imagery enhances the accuracy of biomass estimation and grassland monitoring [13,40]. However, a significant limitation of low-resolution imagery is the difficulty in collecting ground-truth data for large areas, necessitating the use of aggregated or harmonized datasets [9].

Mapping analysis using the spectral harmonized VIIRS (P375) dataset estimated an average biomass of 71.4 g/m² across Mongolia from 2020 to 2024, with notable temporal variability. The highest average biomass occurred in 2024 (78 g/m²), likely due to favorable climatic conditions, while the lowest was in 2022 (60 g/m²), possibly reflecting environmental stress. Biomass estimates varied by grassland type, with desert grasslands showing the largest discrepancy: P375 estimated 54.8 g/m², 10.5% higher than PlanetScope’s 49 g/m². For dry grasslands, P375 estimated 122.6 g/m², 9.6% higher than PlanetScope’s 110.8 g/m². Mountain grasslands exhibited the closest agreement, with P375 estimating 134 g/m², just 1.9% lower than PlanetScope’s 136.6 g/m². Uncertainty analysis further validated these findings, revealing an overall RMSE of 11.6 g/m², a mean percentage difference of 10.74%, and an R² of 0.92 compared to PlanetScope-derived biomass. Mountain grasslands showed the lowest RMSE, though a relatively low R² suggested limited variability, while higher uncertainty in desert and dry grasslands underscored the influence of ecological heterogeneity.

These results highlight the robustness of the spectral harmonized VIIRS dataset for grassland biomass estimations and underscore the importance of considering the grassland type when interpreting the results. In particular, the mountain grassland areas, being more homogenous and predominantly covered by grass, demonstrated better alignment between the two datasets. This observation underscores the critical role of surface type and classification in the adaptation process, emphasizing the need for tailored approaches to improve spectral harmonization accuracy across diverse grassland ecosystems.

Spectral Harmonization Methodology: The SBAF technique was evaluated in optimized and non-optimized forms. The non-optimized approach led to slight increases in discrepancies, while the optimized technique significantly reduced reflectance and NDVI differences. This study uniquely applies SBAF to spectral harmonized PlanetScope and VIIRS imagery for grassland monitoring, achieving notable improvements in grassy and bare areas. This study is the first to spectrally harmonize PlanetScope and VIIRS imagery specifically for grassland applications.

Previous studies have similarly explored harmonization between sensors. For instance, Landsat 7 and MODIS were harmonized using SBAF derived from Hyperion at the Libya 4 calibration site, achieving reductions in discrepancies from 4.26% to 0.93% for the red band and −4.42% to 4.26% for the NIR band [25]. Another study emphasized the importance of surface type in spectral harmonization techniques [26], while SBAFs for multiple sensors have been developed using SCIAMACHY hyperspectral data [27]. Unlike these studies, which used a single calibration site or global coverage, our research focuses on region-specific grasslands and employs an optimized SBAF approach tailored to the surface type.

Potential Limitations: This study focuses exclusively on grasslands and does not propose SBAF for areas with low reflectance, such as those with an NDVI below 0.12 (non-grass areas). Additionally, threshold NDVIs for SBAF optimization was identified based on our method. The reliance on hyperspectral data, such as EO-1 Hyperion, presents challenges due to limited availability and potential data gaps.

The previously developed biomass estimation model was used without modifications for biomass mapping. As a result, the biomass estimation accuracy in this study was constrained by the accuracy of the original model and the additional spatial adjustment errors introduced by spectral harmonization with VIIRS data. Future research could address these limitations by refining the biomass estimation model to account for VIIRS-specific characteristics.

Clear images of homogeneous grassland areas were carefully selected in the harmonization evaluation to ensure consistency rather than using random samples. This selection process may limit the applicability of the findings to more heterogeneous landscapes. Furthermore, the preprocessing of VIIRS data, including converting TOA to BOA reflectance and BRDF corrections, required specialized tools and methodologies. Only nadir observations were used for the evaluation, excluding off-nadir data, which may introduce variability and reduce the general applicability of the results.

Future Directions: Future research will prioritize the classification of the Mongolian area to achieve a detailed understanding of land cover types, enabling more accurate harmonization and modeling. Additionally, advanced techniques, including machine learning and deep learning algorithms, will be developed and applied to enhance the harmonization process and improve modeling accuracy. The ultimate goal is to develop a comprehensive, scalable framework for near real-time grassland ecosystem monitoring, which could serve as a model for other regions facing similar environmental and land-use challenges.

9. Conclusions

This study successfully demonstrated the adaptation of a high-resolution PlanetScope image-driven biomass estimation model to low-resolution VIIRS imagery through spectral harmonizing using the optimized SBAF technique. By addressing discrepancies caused by variations in sensor RSRs, this methodology effectively bridges the resolution gap between high- and low-resolution imagery, providing a robust solution for large-scale grassland monitoring.

The harmonized VIIRS (P375) data estimated an average biomass of 71.4 g/m² across Mongolia from 2020 to 2024, reflecting significant temporal variability. The highest average biomass was recorded in 2024 (78 g/m²), likely due to favorable climatic conditions, while the lowest occurred in 2022 (60 g/m²), potentially indicating environmental stress. Biomass estimates varied by grassland type: desert grasslands showed the largest discrepancy, with P375 estimating 54.8 g/m², 10.5% higher than PlanetScope’s 49 g/m²; dry grasslands yielded 122.6 g/m², 9.6% higher than PlanetScope’s 110.8 g/m²; and mountain grasslands exhibited the closest agreement, with P375 estimating 134 g/m², just 1.9% lower than PlanetScope’s 136.6 g/m². The uncertainty analysis further confirmed the reliability of these estimates, with an overall RMSE of 11.6 g/m², a mean percentage difference of 10.74%, and an R² of 0.92 compared to PlanetScope-derived biomass. These results highlight the effectiveness of spectral harmonized VIIRS data for biomass estimation and emphasize the influence of grassland type on accuracy and interpretation.

The optimized SBAF technique also significantly reduced reflectance discrepancies in both grassy and bare areas. In grassy areas, red band differences decreased from 6.2% to 4.8%, and NDVI discrepancies dropped from 5.5% to 3.1%. For bare areas, red band differences were reduced from 6.9% to 4.0%, and NDVI discrepancies declined from 6.1% to 4.9%. These findings underscore the potential of spectral harmonization to enhance the applicability of high-resolution models to low-resolution datasets, supporting more accurate and frequent monitoring.

With its high temporal resolution and wide swath coverage, VIIRS provides a significant advantage for daily and large-scale monitoring. Its consistent data availability makes it an invaluable tool for applications requiring frequent observations, such as grassland growth tracking, biomass estimation, and drought monitoring. By spectrally harmonizing VIIRS data with high-resolution imagery, this study leveraged these strengths while addressing the limitations of VIIRS’s lower spatial resolution, ensuring its utility for grassland ecosystem monitoring.

While the results are promising, certain limitations remain. These include the reliance on hyperspectral data, predefined threshold NDVIs, and the exclusion of off-nadir observations. Future research should focus on refining the biomass estimation model to better accommodate VIIRS-specific characteristics, integrating additional hyperspectral datasets, and exploring advanced techniques such as machine-learning-based spectral harmonization to further enhance harmonization accuracy.

This study marks a critical step toward developing efficient, scalable solutions for global grassland monitoring in the context of climate change and increasing anthropogenic pressures. By harmonizing high- and low-resolution datasets, the proposed methodology not only improves monitoring capabilities for grasslands but also offers a framework that can be adapted for other ecosystems and regions, contributing to sustainable land management and ecosystem resilience.