Grassland above-ground biomass (AGB) serves as a critical indicator for assessing ecosystem health, carbon sequestration capacity, and sustainable grazing management. In alpine regions, rapid and accurate AGB estimation remains challenging due to complex terrain, high spatial heterogeneity, and limited accessibility. Drone technology, particularly when integrated with hyperspectral sensors, offers a promising solution by providing fine-resolution spectral data that can capture subtle variations in vegetation physiology. This study leverages drone technology to construct a ground–air collaborative inversion framework for AGB estimation in the Gannan Tibetan Autonomous Prefecture, a typical alpine grassland region on the northeastern margin of the Qinghai–Tibet Plateau. By systematically analyzing the effects of flight altitude, phenological timing, and elevation gradient on inversion performance, we demonstrate how drone technology can be optimized to balance spatial detail and operational efficiency. Our results indicate that the XGBoost model integrating multi-order differential features and vegetation indices achieved the highest accuracy (R² = 0.50, RMSE = 12.05 g·m⁻²). Among three tested flight altitudes, 30 m captured fine-scale patchiness well, while 100 m was more suitable for mapping regional trends. Seasonal phenology dominated the temporal dynamics of AGB, and elevation gradients caused spatial patterns to become increasingly fragmented at higher altitudes. These findings confirm the potential of drone technology for precise alpine grassland AGB monitoring and provide practical guidance for scaling drone-based observations across heterogeneous landscapes.
1. Introduction
Accurate estimation of grassland above-ground biomass (AGB) is fundamental for understanding ecosystem function, carbon balance, and livestock carrying capacity. Traditional field sampling methods, while accurate, are labor-intensive, destructive, and impractical for large-scale monitoring. Remote sensing has emerged as an efficient alternative, with multispectral sensors being widely used due to their availability and simplicity. However, multispectral vegetation indices often suffer from saturation effects in high-biomass areas, limiting their accuracy.
Hyperspectral remote sensing, with its contiguous narrow bands, can capture subtle physiological and structural variations in vegetation canopies, offering superior potential for AGB estimation. The integration of drone technology with hyperspectral sensors further enhances this potential by enabling flexible, high-resolution data acquisition at low altitudes. This combination allows researchers to overcome the spatial resolution limitations of satellite sensors and the logistical constraints of ground surveys. Drone technology provides repeated measurements across different flight heights and phenological stages, making it possible to investigate scale effects and temporal dynamics in alpine grasslands.
Despite these advantages, challenges remain in effectively processing high-dimensional hyperspectral data, selecting informative features, and building robust inversion models. Spectral transformation techniques such as fractional-order differentiation have been shown to enhance weak spectral signals and reduce noise, while machine learning algorithms like XGBoost and Random Forest can handle nonlinear relationships. Additionally, the impact of drone flight altitude, phenological timing, and topographic factors on AGB inversion has not been systematically quantified in alpine environments.
In this study, we aim to: (1) develop an optimal AGB inversion model by combining ground-based hyperspectral measurements, drone-derived hyperspectral imagery, and advanced spectral transformations; (2) evaluate the performance of different flight altitudes (30 m, 70 m, 100 m) in capturing spatial heterogeneity; (3) analyze the seasonal dynamics of AGB across key phenological stages; and (4) investigate the effect of elevation gradients on AGB spatial patterns. By addressing these objectives, we demonstrate how drone technology can be harnessed to deliver accurate and scalable AGB estimates for alpine grassland management.
2. Materials and Methods
2.1 Study Area
The study was conducted in Gannan Tibetan Autonomous Prefecture, located in the northeastern margin of the Qinghai–Tibet Plateau (100°46′–104°44′E, 33°06′–36°10′N). The area covers approximately 45,000 km² with elevations ranging from 1,100 m to 4,900 m. Alpine meadows dominate the vegetation, covering over 90% of the grassland area. The climate is cold and semi-humid, with precipitation concentrated from June to September. We selected five representative counties (Maqu, Xiahe, Hezuo, Zhuoni, and Luqu) for field sampling, covering a range of elevations (1,516–3,928 m) and environmental conditions.
2.2 Data Acquisition
2.2.1 Ground-Based Hyperspectral and AGB Data
We collected 375 samples from 49 plots between May and September 2019. In each plot, three 1 m × 1 m quadrats were harvested along 0°, 120°, and 240° directions from the center. All above-ground plant material was oven-dried at 65 °C to constant weight and weighed to obtain AGB in g·m⁻². Concurrently, canopy hyperspectral reflectance was measured using an ASD FieldSpec HandHeld 4 spectroradiometer covering 350–2,500 nm with a spectral resolution of 3 nm. Each spectral measurement was averaged over multiple scans and recorded along with GPS coordinates to ensure accurate spatial matching with drone data.
2.2.2 Drone Hyperspectral Data
Drone technology was employed using a DJI Matrice 600 Pro hexacopter equipped with a GaiaSky-omo2-VN visible/near-infrared hyperspectral imaging system (400–1,000 nm). Flight missions were conducted during the peak growing season (June–September) at three altitudes: 30 m, 70 m, and 100 m. A reflectance panel was placed at the center of each flight area for radiometric calibration. Raw images underwent radiometric, atmospheric, and geometric corrections. To align ground and drone spectra, we used GPS coordinates to extract pixel values from the corrected drone imagery corresponding to each ground quadrat.
2.3 Spectral Transformation and Feature Extraction
To enhance spectral sensitivity to AGB, we applied several transformations to the raw reflectance data:
- Fractional-order differentiation: Using the Grunwald–Letnikov definition, we computed derivatives from order 0 to 1 with a step of 0.2, as well as second-order derivatives (order 2.0). The general formula for fractional-order differentiation of order α is:
$$D^{\alpha}f(x) = \lim_{h \to 0} \frac{1}{h^\alpha} \sum_{k=0}^{\lfloor (x-a)/h \rfloor} (-1)^k \binom{\alpha}{k} f(x – kh)$$
where α is the order and h is the step size. We implemented this numerically with h = 1 nm.
- Logarithmic transformation: Pseudo-absorption coefficient A = log(1/R) was computed to linearize the relationship with absorbing constituents.
- Vegetation indices: A suite of 20 vegetation indices was calculated, including NDVI, CIred-edge, SAVI, OSAVI, GNDVI, TCARI, MCARI, MTCI, PRI, SIPI, and others (Table 1).
Table 1. Vegetation indices used in this study.
| Index | Full Name | Formula |
|---|---|---|
| NDVI | Normalized Difference Vegetation Index | (NIR − RED) / (NIR + RED) |
| CIred-edge | Chlorophyll Index (Red Edge) | (NIR / RedEdge) − 1 |
| GNDVI | Green NDVI | (NIR − GREEN) / (NIR + GREEN) |
| SAVI | Soil Adjusted Vegetation Index | ((NIR − RED) / (NIR + RED + L)) × (1+L), L=0.5 |
| OSAVI | Optimized SAVI | (NIR − RED) / (NIR + RED + 0.16) × (1.16) |
| CARI | Chlorophyll Absorption Ratio Index | ((a × RED) + b) / √(a²+1), a = (R700 − R550)/150 |
| TCARI | Transformed Chlorophyll Absorption in Reflectance Index | 3 × [(R700 − R670) − 0.2 × (R700 − R550) × (R700 / R670)] |
| MCARI | Modified Chlorophyll Absorption in Reflectance Index | [(R700 − R670) − 0.2 × (R700 − R550)] × (R700 / R670) |
| HNDVI | Hyperspectral NDVI | (R750 − R705) / (R750 + R705) |
| MTCI | MERIS Terrestrial Chlorophyll Index | (R754 − R709) / (R709 − R681) |
| PRI | Photochemical Reflectance Index | (R531 − R570) / (R531 + R570) |
| SIPI | Structure Insensitive Pigment Index | (R800 − R445) / (R800 + R680) |
| PSNDa | Pigment Specific NDVI a | (NIR − R680) / (NIR + R680) |
| PSNDb | Pigment Specific NDVI b | (NIR − R635) / (NIR + R635) |
| PSSRa | Pigment Specific Simple Ratio a | NIR / R680 |
| PSSRb | Pigment Specific Simple Ratio b | NIR / R635 |
| VARIg | Visible Atmospherically Resistant Index (green) | (GREEN − RED) / (GREEN + RED − BLUE) |
| VARIr | VARI using red-edge | (RedEdge − RED) / (RedEdge + RED − BLUE) |
| SR | Simple Ratio | NIR / RED |
| TVI | Transformed Vegetation Index | √(NDVI + 0.5) |
| GRVI | Green-Red Vegetation Index | (GREEN − RED) / (GREEN + RED) |
2.4 Feature Selection
We employed a two-stage feature selection strategy. First, correlation analysis (Pearson’s r) between each transformed spectral feature and AGB was performed. Second, we applied the Random Frog algorithm based on Random Forest (RF-Random Frog) to identify the most relevant feature subset. This method iteratively explores the feature space by simulating frog jumping behavior and evaluates candidate subsets using Random Forest’s variable importance. The top 10 features from each transformation were retained, and then combined with vegetation indices to form the final feature set (Table 2).
Table 2. Selected spectral bands and combined features.
| Dataset | Selected bands (nm) |
|---|---|
| Original spectrum | 385, 379, 434, 848, 380, 701, 925, 735, 830, 629 |
| 0.2-order derivative | 384, 382, 694, 441, 699, 747, 815, 917, 707 |
| 0.4-order derivative | 386, 369, 909, 439, 619, 682, 373, 838, 782, 926 |
| 0.6-order derivative | 369, 390, 645, 412 |
| 0.8-order derivative | 354, 397, 412, 626, 455, 608, 776, 572, 925, 828 |
| 1.0-order derivative | 626, 793, 404, 688, 747, 420, 632, 762, 425, 776 |
| 2.0-order derivative | 883, 574, 431, 592, 948, 921, 571, 825, 820, 990 |
| Logarithmic transform | 360, 460, 381, 591, 700, 754, 787, 517, 601, 733 |
| Bands + Vegetation indices | log_460, SAVI, 0.4_386, log_512, CIred-edge, 1.0_626, 0.8_354, 0.6_369, SR, 0.6_390 |
2.5 Modelling Framework
We built AGB inversion models using two machine learning algorithms: XGBoost (XGB) and Random Forest (RF). For XGBoost, hyperparameters were optimized using Bayesian optimization (Optuna) over 50 iterations, with learning rate in [0.01, 0.3], tree depth in [3, 10], and minimum child weight in [1, 10]. Early stopping was applied after 20 rounds without improvement. For Random Forest, the number of trees was set to 500, and the number of features considered at each split was tuned via out-of-bag error. Five-fold cross-validation was used to evaluate model performance.
Model accuracy was assessed using the coefficient of determination (R²) and 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 yi is the measured AGB, ŷi is the predicted AGB, and ȳ is the mean measured AGB.
3. Results
3.1 Spectral Response Patterns
Correlation analysis between different spectral transformations and AGB revealed wavelength-dependent patterns (Fig. 2 in original study). In the visible region (400–600 nm), original reflectance and low-order fractional derivatives showed strong positive correlations with AGB, particularly in the blue and green regions. The red-edge region (680–750 nm) exhibited a shift from positive to negative correlation as derivative order increased. Near-infrared bands (800–950 nm) consistently showed negative correlations, likely due to increased canopy shadowing at higher biomass levels. The selected features from Random Frog (Table 2) confirmed that bands in the blue (380–460 nm), red-edge (680–747 nm), and near-infrared (810–925 nm) were most informative.
3.2 Model Performance and Feature Selection
We compared the performance of XGBoost and Random Forest across different spectral transformations and feature sets (Table 3). XGBoost consistently outperformed Random Forest in both R² and RMSE, especially when using higher-order derivatives and combined features. The best-performing model was XGBoost trained on the combined feature set of “bands summary + vegetation indices” (46 variables), achieving R² = 0.50 and RMSE = 12.05 g·m⁻². Among single transformations, the 1.0-order derivative gave the best results (R² = 0.47, RMSE = 13.56 g·m⁻²), outperforming the original spectrum (R² = 0.41, RMSE = 10.27 g·m⁻²). The 2.0-order derivative performed poorly (R² = 0.27, RMSE = 19.90 g·m⁻²), indicating excessive noise amplification.
Table 3. Model performance for different spectral transformations and feature sets.
| Dataset | Model | Number of features | R² | RMSE (g·m⁻²) |
|---|---|---|---|---|
| 0.2-order derivative | RF | 9 | 0.42 | 38.12 |
| 0.4-order derivative | XGBoost | 14 | 0.40 | 22.52 |
| 0.6-order derivative | RF | 4 | 0.43 | 25.29 |
| 0.8-order derivative | RF | 11 | 0.39 | 31.80 |
| 1.0-order derivative | XGBoost | 14 | 0.47 | 13.56 |
| 2.0-order derivative | XGBoost | 17 | 0.27 | 19.90 |
| Original spectrum | XGBoost | 12 | 0.41 | 10.27 |
| Logarithmic transform | XGBoost | 14 | 0.42 | 14.55 |
| Bands summary | XGBoost | 97 | 0.48 | 26.74 |
| Bands summary + VIs | XGBoost | 46 | 0.50 | 12.05 |
3.3 Effect of Drone Flight Altitude on AGB Inversion
We applied the optimal XGBoost model to drone hyperspectral imagery acquired at 30 m, 70 m, and 100 m altitudes over two representative sites (Guomang and Dashui). The spatial patterns of estimated AGB are shown in Figure 3 of the original study. At 30 m, the high spatial resolution captured fine-scale patchiness, including distinct boundaries between grass patches and bare soil, but also introduced noise from individual shadows and gaps. At 70 m, the inversion results were smoother and more continuous, effectively balancing detail and stability. At 100 m, although spatial details were reduced, the overall biomass gradient was well preserved, making this altitude suitable for broader regional mapping. Quantitative comparison (Table 4) shows that RMSE increased slightly with altitude, but the 70 m flight provided the best compromise between resolution and noise suppression.
Table 4. Comparison of AGB inversion metrics at different drone flight altitudes.
| Flight altitude | Mean estimated AGB (g·m⁻²) | Relative RMSE (%) | Spatial detail |
|---|---|---|---|
| 30 m | 68.2 | 18.4 | High (patchy) |
| 70 m | 65.7 | 15.2 | Moderate (smooth) |
| 100 m | 64.1 | 16.8 | Low (regional trend) |
3.4 Phenological Dynamics of AGB
Drone technology enabled repeated monitoring of three typical sites (Azi, Dashui, and Yaliji) in June, July, and September 2019 (growing season). The AGB maps revealed distinct temporal patterns (Fig. 4 in original study). Azi exhibited a unimodal pattern: low in June, peak in July, and decline in September, consistent with rapid growth and senescence of alpine meadow species. Dashui showed a more complex pattern, with relatively high AGB in June and September but a dip in July, possibly due to grazing pressure or water stress. Yaliji displayed a continuous increase from June to September, accumulating biomass until late season, but with increasing spatial heterogeneity. These results highlight that drone technology can capture site-specific phenological variability, which is essential for accurate AGB monitoring.
3.5 Elevation Gradient Effect
We selected three elevation zones (3,100 m, 3,300 m, and 3,500 m) within the study area and applied the inversion model to drone imagery. The results (Fig. 5 in original study) showed a clear decreasing trend in AGB with increasing elevation. At 3,100 m, AGB was high and uniformly distributed. At 3,300 m, biomass remained moderately high but began to show patchiness. At 3,500 m, AGB was significantly lower, with many low-biomass patches. This pattern is attributed to decreasing temperature, shorter growing season, and more fragmented vegetation cover at higher elevations. The ability of drone technology to resolve such fine-scale elevation gradients demonstrates its value for topographic ecology studies.
4. Discussion
Our study demonstrates that drone technology, when combined with advanced spectral transformations and machine learning, can provide accurate and spatially explicit estimates of alpine grassland AGB. The best model (XGBoost with combined features) achieved an R² of 0.50, which is competitive with similar studies using satellite or airborne data, but with the added advantage of flexible revisit times and ultra-high resolution. The unexplained variance (50%) likely stems from residual soil background effects, species composition differences, and scale mismatches between ground quadrats and drone pixels. Future integration of structural metrics (e.g., from LiDAR) or multitemporal sequences could further improve accuracy.
Drone flight altitude emerged as a critical factor. The 30 m altitude captured fine-scale heterogeneity well but introduced noise; the 70 m altitude offered the best balance; and the 100 m altitude was effective for regional trends. This finding provides practical guidance for drone technology deployment: high-resolution flights for local patch studies and lower-resolution flights for extensive surveys. Similarly, phenological timing strongly influenced AGB estimates; single-date snapshots may be misleading if not aligned with peak biomass. Drone technology’s ability to conduct repeated flights at low cost makes it ideal for capturing seasonal dynamics.
The observed elevation gradient (decreasing AGB with altitude) aligns with known thermal limitations and precipitation patterns in alpine environments. Drone technology allowed us to map these gradients at a spatial resolution that satellite sensors cannot achieve, revealing the fragmentation of vegetation at high elevations.
Nevertheless, our study has limitations. The sample size (375) is moderate and spatially clustered, potentially reducing model generalizability. The models were calibrated only for the Gannan region; transferability to other alpine areas needs validation. Additionally, our feature set was purely spectral; including canopy height or soil moisture data could improve performance. Future work should expand sampling across different grassland types, incorporate multitemporal drone flights throughout the entire growing season, and explore deep learning approaches that can learn hierarchical features directly from hyperspectral cubes.
5. Conclusion
This study successfully developed a ground–air collaborative AGB inversion framework for alpine grasslands using drone technology and hyperspectral data. The key contributions are: (1) Spectral transformations, particularly 1.0-order fractional derivatives combined with vegetation indices, significantly enhanced model accuracy. (2) XGBoost outperformed Random Forest, achieving R² = 0.50 and RMSE = 12.05 g·m⁻². (3) Drone flight altitude affects the trade-off between spatial detail and noise; 70 m is recommended for balanced performance. (4) Phenological timing and elevation gradient strongly modulate AGB spatial patterns, and drone technology enables their systematic characterization. Our results underscore the potential of drone technology as an operational tool for alpine grassland monitoring, supporting sustainable pasture management and ecological conservation in sensitive high-altitude regions.