The quantification of Above-Ground Biomass (AGB) in grasslands is fundamental for assessing ecosystem health, carbon sequestration potential, and sustainable management of pastoral resources. Traditional ground-based survey methods, while accurate, are destructive, labor-intensive, and lack the scalability required for dynamic, large-area monitoring. Remote sensing offers a viable alternative, with multispectral data from satellites and airborne platforms having been extensively applied. However, these approaches often face limitations, such as spectral band saturation in dense vegetation and coarse spatial resolution that fails to capture fine-scale heterogeneity. The emergence of unmanned drone technology equipped with lightweight hyperspectral sensors presents a transformative opportunity. These systems bridge the gap between detailed ground measurements and broad-scale satellite observations, offering high spectral fidelity, very high spatial resolution, and unparalleled operational flexibility.
This study explores the potential of integrating ground-based and unmanned drone-borne hyperspectral data for the fine-scale estimation of AGB in alpine grasslands. Alpine ecosystems, such as those on the northeastern fringe of the Tibetan Plateau, are highly sensitive to climate change and human activities, making accurate biomass monitoring crucial. We developed a synergistic ground-unmanned aerial vehicle inversion framework, employing advanced spectral transformations and machine learning algorithms. Furthermore, we systematically analyzed how key operational and environmental factors—specifically unmanned drone flight altitude, seasonal phenology, and elevation gradient—influence the inversion results and their spatial expression.
1. Materials and Methods
1.1 Study Area
The study was conducted in the Gannan Tibetan Autonomous Prefecture, located on the transitional zone between the northeastern Tibetan Plateau and the Loess Plateau. This region serves as a critical national ecological barrier. The terrain exhibits significant elevational gradients, ranging from approximately 1,100 to over 4,900 meters, with the majority of the area above 3,000 meters, characterizing a typical high-altitude, cold environment. The climate is temperate semi-humid to humid, with precipitation concentrated from June to September. Alpine meadow is the dominant grassland type, accounting for over 90% of the total grassland area, primarily distributed in the western and southern parts of the prefecture. These meadows are vital for local animal husbandry and ecological conservation.
1.2 Data Acquisition and Processing
1.2.1 Ground-Measured AGB and Spectral Data
A total of 375 AGB samples were collected from 49 representative plots across the prefecture during the growing season (May to September) of 2019. Sampling followed a standardized protocol: three 1 m × 1 m quadrats were established at 40 m intervals from a central GPS point. All above-ground vegetation within each quadrat was clipped, oven-dried at 65°C to constant weight, and weighed to determine AGB in grams per square meter (g·m⁻²). Concurrently, canopy spectral reflectance was measured for each quadrat using an Analytical Spectral Devices (ASD) FieldSpec HandHeld 4 spectroradiometer (350–2500 nm range, 1.4 nm sampling interval). Precise GPS coordinates and timestamps were recorded to ensure accurate co-registration with unmanned drone imagery.
1.2.2 Unmanned Drone Hyperspectral Data
Hyperspectral image data were acquired using a DJI Matrice 600 Pro unmanned aerial vehicle equipped with a GaiaSky-omo2-VN hyperspectral imaging system (400–1000 nm). Flights were conducted at three different altitudes (30 m, 70 m, and 100 m) across key phenological stages from June to September. A calibrated reflectance panel was deployed within each flight area for radiometric correction to mitigate the effects of varying illumination conditions. The raw imagery underwent standard preprocessing, including radiometric calibration, atmospheric correction, and geometric correction using ground control points and high-precision GPS data.
1.2.3 Spectral Feature Engineering
To enhance the sensitivity of spectral data to AGB and mitigate issues of high dimensionality and noise, several mathematical transformations were applied to both ground and unmanned drone spectral data:
- Fractional-Order Differentiation (FOD): The Grunwald–Letnikov definition was used to calculate fractional derivatives from order 0 to 1 with a step of 0.2, and the second-order derivative was also computed. This method extracts multi-scale spectral variation information. The FOD of a signal \( R(\lambda) \) is defined as:
$$ D^{\nu}[R(\lambda)] = \lim_{h \to 0} h^{-\nu} \sum_{m=0}^{\lfloor (\lambda-a)/h \rfloor} (-1)^m \binom{\nu}{m} R(\lambda – mh) $$
where \( \nu \) is the fractional order, \( h \) is the step, and \( a \) is the lower limit. - Logarithmic Transformation (Log(1/R)): Calculated to approximate absorption features based on the Beer-Lambert law.
$$ \text{Absorption} = \log_{10}(1 / R_{\lambda}) $$ - Vegetation Indices (VIs): A suite of established VIs was calculated to integrate information from key spectral regions related to chlorophyll content, canopy structure, and soil background adjustment. Key indices are listed in Table 1.
| Vegetation Index | Full Name | Formula |
|---|---|---|
| NDVI | Normalized Difference Vegetation Index | $$ (R_{NIR} – R_{Red}) / (R_{NIR} + R_{Red}) $$ |
| SAVI | Soil Adjusted Vegetation Index | $$ ((R_{NIR} – R_{Red}) / (R_{NIR} + R_{Red} + L)) \times (1 + L), L=0.5 $$ |
| OSAVI | Optimized Soil Adjusted Vegetation Index | $$ (R_{NIR} – R_{Red}) / (R_{NIR} + R_{Red} + 0.16) $$ |
| CIred-edge | Chlorophyll Index – Red Edge | $$ (R_{NIR} / R_{RedEdge}) – 1 $$ |
| MCARI | Modified Chlorophyll Absorption Ratio Index | $$ [(R_{700} – R_{670}) – 0.2 \times (R_{700} – R_{550})] \times (R_{700} / R_{670}) $$ |
| PRI | Photochemical Reflectance Index | $$ (R_{531} – R_{570}) / (R_{531} + R_{570}) $$ |
1.3 Feature Selection and Model Development
A multi-strategy feature selection approach was employed to identify the most predictive variables from the high-dimensional feature space, which included original bands, FOD-transformed data, and VIs. First, correlation analysis and Principal Component Analysis (PCA) were used for an initial screening of bands highly correlated with AGB. Subsequently, the Random Frog algorithm, optimized with a Random Forest (RF) learner (RF-Random Frog), was applied to the candidate feature set to search for the optimal variable subset, enhancing the ability to handle nonlinear relationships.
Prior to model integration, the spectral consistency between ground ASD measurements and unmanned drone image-extracted reflectance for the same sample plots was rigorously assessed using Bland-Altman analysis and Pearson correlation. This confirmed the data were suitable for synergistic modeling.
Two powerful ensemble learning algorithms, eXtreme Gradient Boosting (XGBoost) and Random Forest (RF), were used to construct the AGB estimation models. For the XGBoost model, the Optuna framework with Bayesian optimization was used for hyperparameter tuning (e.g., learning rate, tree depth). Early stopping was implemented to prevent overfitting. The RF model used out-of-bag error for evaluation and Gini importance for feature ranking. Model training and evaluation were performed using 5-fold cross-validation.
Model performance was evaluated using the Coefficient of Determination (\(R^2\)) and the Root Mean Square Error (RMSE):
$$ R^2 = 1 – \frac{\sum_{i=1}^{n} (y_i – \hat{y}_i)^2}{\sum_{i=1}^{n} (y_i – \bar{y})^2} $$
$$ RMSE = \sqrt{\frac{1}{n} \sum_{i=1}^{n} (y_i – \hat{y}_i)^2} $$
where \(y_i\) is the measured AGB, \(\hat{y}_i\) is the predicted AGB, \(\bar{y}\) is the mean of measured AGB, and \(n\) is the number of samples.
2. Results
2.1 Hyperspectral Feature Analysis
Correlation heatmaps revealed distinct spectral response patterns across different transformations. In the visible region (400–600 nm), most transformations showed positive correlations with AGB, particularly for original reflectance and lower-order (0.2–0.6) FOD, indicating sensitivity to chlorophyll absorption. The red-edge region (680–750 nm) showed a reversal: correlations were positive for low-order FOD and log transformation but became strongly negative for first and second derivatives, likely due to saturation effects and slope variations. The near-infrared region (800–950 nm) predominantly showed negative correlations across transformations.
The RF-Random Frog algorithm selected key sensitive bands from different transformations, as summarized in Table 2. Bands in the blue-violet region (~380–460 nm, related to chlorophyll absorption) and specific regions in the red-edge and near-infrared (~680–950 nm, related to canopy structure and moisture) were consistently selected across multiple fractional orders and the logarithmic transformation.
| Feature Set | Selected Bands / Features (Example) |
|---|---|
| Original Spectrum | 385, 379, 434, 848, 925 nm |
| 0.4-order FOD | 386, 369, 909, 682 nm |
| 1.0-order FOD | 626, 688, 747, 776 nm |
| Log(1/R) | 460, 700, 754, 787 nm |
| Fused Set (Bands + VIs) | log_460, SAVI, 0.4_386, CIred-edge, 1.0_626, SR |
2.2 AGB Model Inversion and Accuracy
The consistency analysis between ground and unmanned drone spectra showed high agreement, with Pearson correlation coefficients exceeding 0.98 for all sample groups, validating the basis for synergistic modeling.
The performance of AGB inversion models varied with the spectral preprocessing method and algorithm (Table 3). The XGBoost model generally outperformed the RF model in terms of stability and predictive accuracy. Among the fractional-order differentials, the 1.0-order transformation achieved the best balance between feature enhancement and noise suppression, yielding an \(R^2\) of 0.47 with XGBoost. The highest accuracy was achieved by the XGBoost model trained on the fused feature set combining key bands from multiple transformations and vegetation indices (e.g., SAVI, CIred-edge), resulting in an \(R^2\) of 0.50 and an RMSE of 12.05 g·m⁻².
| Feature Set | Model | No. of Features | R² | RMSE (g·m⁻²) |
|---|---|---|---|---|
| Original Spectrum | XGBoost | 12 | 0.41 | 10.27 |
| 1.0-order FOD | XGBoost | 14 | 0.47 | 13.56 |
| Log(1/R) | XGBoost | 14 | 0.42 | 14.55 |
| Band Summary | XGBoost | 97 | 0.48 | 26.74 |
| Bands + Vegetation Indices | XGBoost | 46 | 0.50 | 12.05 |
2.3 Impact of Unmanned Drone Flight Altitude
The flight altitude of the unmanned drone significantly influenced the spatial detail and stability of the AGB inversion maps. Imagery from the 30 m flight altitude provided the highest spatial resolution, effectively capturing fine-scale heterogeneity and patchiness within grassland plots. However, it was also more susceptible to noise from micro-variations and shadows. The 70 m altitude offered a good compromise, maintaining considerable spatial detail while producing more continuous and stable inversion results. The 100 m altitude maps showed reduced fine detail but reliably captured the overall spatial pattern and trends of AGB distribution over larger areas, demonstrating its utility for regional-scale assessment.
2.4 Impact of Phenological Timing
The inversion of AGB across different months (June, July, September) revealed distinct seasonal dynamics. Some sample plots exhibited a classic single-peak pattern, with low biomass in June, a peak in July, and a decline by September, aligning with the typical growth cycle of alpine grasses. Other plots showed more complex temporal patterns, such as a mid-season dip or a continuous increase until September. These differences underscore that local factors like moisture availability, grazing pressure, or species composition can modulate the phenological signal captured by the unmanned drone, leading to significant spatiotemporal heterogeneity in AGB.
2.5 Impact of Elevation Gradient
A clear altitudinal trend was observed in the spatial patterns of estimated AGB. In lower elevation zones (e.g., ~3100 m), AGB was generally higher and more uniformly distributed. At middle elevations (~3300 m), AGB remained relatively high but began to show a more patchy distribution. At the highest elevations examined (~3500 m), the overall AGB level decreased noticeably, and its spatial distribution became significantly more fragmented, with low-biomass patches dominating the landscape. This pattern is consistent with the increasing environmental stress (e.g., lower temperatures, shorter growing seasons) associated with higher altitudes in alpine regions.
3. Discussion
This study demonstrates the significant potential of unmanned drone hyperspectral remote sensing for fine-scale AGB estimation in ecologically sensitive alpine grasslands. The high spectral consistency between ground and unmanned drone measurements validated the core premise of our synergistic modeling approach. The integration of advanced spectral processing techniques, particularly fractional-order differentiation, proved effective in enhancing the feature space and mitigating the saturation effects common in high-coverage vegetation. The fusion of these mathematically enhanced spectral features with biophysically meaningful vegetation indices (like SAVI and CIred-edge) within a powerful nonlinear learner (XGBoost) yielded the most robust model. This highlights that moving beyond simple vegetation indices or raw spectra to a multi-feature, machine-learning framework is key to improving unmanned drone-based biomass estimation.
Our analysis of operational and environmental factors provides practical insights for survey design. The choice of unmanned drone flight altitude involves a trade-off between spatial detail and mapping stability/coverage, which should be guided by the specific monitoring objective. The observed phenological and elevational variations in AGB patterns emphasize that unmanned drone campaigns should be carefully timed and stratified to capture the dominant sources of variability in the ecosystem. The unexplained variance (~50%) in our best model indicates that AGB is influenced by factors beyond spectral reflectance alone, such as community structure, soil properties, and topography. Future work should integrate structural data from LiDAR or radar sensors, multi-temporal unmanned drone observations, and ancillary environmental variables to build more comprehensive and mechanistically informed models.
Limitations of this study include the constrained sample size and spatial distribution, which may affect model generalizability. The analysis was also based on snapshots from key phenological stages rather than a continuous time series. Future research should expand sample coverage across broader environmental gradients and employ time-series unmanned drone data to model seasonal and inter-annual AGB dynamics.
4. Conclusion
This research established a viable ground-unmanned aerial vehicle synergistic framework for estimating alpine grassland AGB using hyperspectral data. The optimal model, which integrated multi-order fractional differential features with selected vegetation indices using the XGBoost algorithm, achieved satisfactory accuracy (\(R^2 = 0.50\), RMSE = 12.05 g·m⁻²). We found that unmanned drone flight altitude significantly affects the spatial expression of inversion results, with lower altitudes (30 m) suitable for fine-scale heterogeneity and higher altitudes (100 m) appropriate for regional trend analysis. Furthermore, seasonal phenology and elevation gradients were key drivers of the spatiotemporal patterns in estimated AGB, with biomass generally decreasing and becoming more patchy at higher elevations. The findings confirm the value of unmanned drone hyperspectral technology for high-resolution grassland monitoring and provide guidance for optimizing remote sensing scale and interpretation in complex alpine environments. To enhance model generality and ecological interpretability, future efforts should focus on expanding sample representation, incorporating multi-source data fusion, and developing time-series monitoring capabilities.