Li et al. (2026) Multidimensional Validation of FVC Products over Qinghai–Tibetan Plateau Alpine Grasslands: Integrating Spatial Representativeness Metrics with Machine Learning Optimization

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-01-10
• Authors: Junji Li, J. L. Chen, Xue Cheng, JIAYUAN YIN, Qingmin Cheng, Haotian You, Xiaowen Han, Xiaodong Li
• DOI: 10.3390/rs18020228

Research Groups

• Guangxi Key Laboratory of Spatial Information and Geomatics, College of Geomatics and Geoinformation, Guilin University of Technology, Guilin 541004, China
• Guizhou Institution of Prataculture, Guizhou Academy of Agricultural Sciences, Guiyang 550006, China

Short Summary

This study developed a comprehensive framework integrating spatial representativeness metrics and machine learning optimization to systematically assess the accuracy of GEOV3, GLASS, and MuSyQ Fractional Vegetation Cover (FVC) products over Qinghai–Tibetan Plateau alpine grasslands. The framework significantly enhanced validation reliability, revealing GEOV3's superior accuracy compared to GLASS and MuSyQ, which consistently underestimated FVC.

Objective

• To develop and apply representativeness metrics, namely the Relative Spatial Sampling Error (RSSE) and an NDVI-based Spatial Heterogeneity Index (NDVI-SHI), to optimize the suitability of ground measurements for FVC validation.
• To comparatively analyze the performance of Random Forest (RF), Gradient Boosting Machine (GBM), Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) models for FVC retrieval and identify the optimal algorithm for generating reference validation datasets.
• To produce a high-spatial-resolution validation dataset and establish a multidimensional validation framework to evaluate the applicability of GEOV3, GLASS, and MuSyQ FVC products over alpine grasslands on the Qinghai–Tibetan Plateau (QTP), thereby providing a robust foundation for product utilization.

Study Configuration

• Spatial Scale: Qinghai–Tibetan Plateau (QTP), approximately 2.57 × 10^6 square kilometers. UAV imagery at approximately 2 centimeter resolution, 30 meter × 30 meter subplots, 250 meter × 250 meter plots. Satellite data at 10 meter (Sentinel-2), 250 meter (MODIS), 300 meter (GEOV3), 500 meter (GLASS, MuSyQ). Topographic data at 30 meter (SRTM DEM). Climatic data at 0.1 degree spatial resolution (ERA5).
• Temporal Scale: July–August periods spanning 2016–2019 for UAV and satellite data collection and compositing. ERA5 reanalysis data available since 1979. GEOV3 FVC product spans 2014 to present, GLASS FVC product spans 2000–2020.

Methodology and Data

• Models used:
  • Machine Learning: Random Forest (RF), Gradient Boosting Machine (GBM), Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost).
  • Representativeness Metrics: Spatial Sampling Error (SSE), Relative Spatial Sampling Error (RSSE), NDVI-based Spatial Heterogeneity Index (NDVI-SHI).
  • Optimization Techniques: Recursive Feature Elimination (RFE), Grid Search with five-fold cross-validation.
• Data sources:
  • Measured data: FVC ground data (UAV imagery at ~2 cm resolution) collected from 623 plots across the QTP.
  • Satellite data: Sentinel-2 Multispectral Instrument (10 m, 20 m spectral bands), MODIS MOD13Q1 vegetation index product (NDVI, EVI at 250 m).
  • FVC product data: Copernicus GEOV FCover (GEOV3, 300 m), GLASS FVC (500 m), MuSyQ FVC (500 m).
  • Auxiliary data: Shuttle Radar Topography Mission Digital Elevation Model (SRTM DEM, 30 m), ERA5 reanalysis (climatic variables at 0.1°).
  • Platform: Google Earth Engine (GEE).

Main Results

• A comprehensive framework, integrating spatial representativeness metrics (RSSE < 0.20 and NDVI-SHI < 0.15) and machine learning optimization, significantly enhanced the robustness of FVC product validation over alpine grasslands.
• CatBoost achieved the superior performance for generating the 10 meter FVC reference dataset (R^2 = 0.880, RMSE = 0.122), followed by XGBoost (R^2 = 0.879, RMSE = 0.123), GBM (R^2 = 0.875, RMSE = 0.125), LightGBM (R^2 = 0.870, RMSE = 0.127), and RF (R^2 = 0.867, RMSE = 0.129).
• GEOV3 consistently demonstrated superior accuracy across all four validation scenarios compared to GLASS and MuSyQ FVC products on alpine grasslands.
• In the most robust validation setting (Scenario 4: multi-scale validation with representativeness-screened 10 m FVC reference), GEOV3 achieved R^2 values ranging from 0.909–0.925 and RMSE values from 0.082–0.103.
• Under the same robust setting, GLASS achieved R^2 values from 0.742–0.771 and RMSE values from 0.138–0.175, while MuSyQ achieved R^2 values from 0.739–0.746 and RMSE values from 0.138–0.181. Both GLASS and MuSyQ exhibited systematic underestimation.
• Representativeness screening (RSSE and NDVI-SHI) effectively reduced mismatch-related errors, improving GEOV3's R^2 from 0.801–0.843 to 0.879–0.900 and decreasing RMSE from 0.122–0.140 to 0.094–0.118 in direct validation.

Contributions

• Developed a novel multidimensional FVC product evaluation framework that integrates spatial representativeness screening, machine learning-based reference construction, and scenario-based validation across scales, specifically designed to address challenges in heterogeneous alpine environments.
• Introduced and validated the effectiveness of combining Relative Spatial Sampling Error (RSSE) and NDVI-based Spatial Heterogeneity Index (NDVI-SHI) for screening ground measurements, significantly reducing validation uncertainty caused by point-to-pixel mismatch and surface heterogeneity.
• Identified and optimized machine learning models, with CatBoost demonstrating superior performance, for generating high-precision, high-spatial-resolution (10 meter) FVC reference datasets from Sentinel-2 imagery over the Qinghai–Tibetan Plateau.
• Provided a robust, comparative accuracy assessment of GEOV3, GLASS, and MuSyQ FVC products over QTP alpine grasslands, offering critical scientific guidance for selecting appropriate FVC products and improving retrieval algorithms for alpine ecosystem monitoring.
• The proposed framework offers a methodological basis for optimizing field sampling design and achieving scale-matched validation of satellite FVC products in highly heterogeneous landscapes.

Funding

• National Natural Science Foundation of China (42361023)
• Natural Science Foundation of Guangxi Province (2024GXNSFAA010268)
• Guangxi Key Laboratory of Spatial Information and Geomatics (21-238-21-31)
• BaGui Young Outstanding Talents Cultivation Program of the provincial government of Guangxi

Citation

@article{Li2026Multidimensional,
  author = {Li, Junji and Chen, J. L. and Cheng, Xue and YIN, JIAYUAN and Cheng, Qingmin and You, Haotian and Han, Xiaowen and Li, Xiaodong},
  title = {Multidimensional Validation of FVC Products over Qinghai–Tibetan Plateau Alpine Grasslands: Integrating Spatial Representativeness Metrics with Machine Learning Optimization},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18020228},
  url = {https://doi.org/10.3390/rs18020228}
}