TL;DR
This paper presents a novel approach to accurately estimate vegetation biomass along expressways using UAV-based LiDAR and multispectral data combined with a Random Forest model. By creating a comprehensive sample library and employing advanced spatial connection techniques, the study achieves high prediction accuracy, offering a robust framework for ecological monitoring and sustainable transportation development.
What Problem Does This Paper Address?
Accurately assessing vegetation biomass is crucial for carbon stock management and ecological restoration, especially in the context of "Carbon Neutrality" goals. Traditional field surveys for biomass estimation are time-consuming, labor-intensive, and limited in scope. While remote sensing offers a solution, existing methods often struggle with spatial resolution, revisit cycles, and the ability to capture 3D structural information, particularly in complex environments like expressway corridors. This paper tackles the challenge of precisely and efficiently monitoring biomass in these dynamic roadside ecosystems, which are vital for soil stabilization and ecological balance.
Methods
The study integrates data from Unmanned Aerial Vehicles (UAVs) equipped with both LiDAR and multispectral sensors. LiDAR data provides detailed 3D structural information (plant height, crown width, DBH), while multispectral data captures spectral features related to vegetation health and type.
Key methodological steps include:
- Data Acquisition: UAVs collected LiDAR point clouds and multispectral imagery over expressway corridors in Shanxi Province.
- Data Preprocessing: Standardized preprocessing of LiDAR (coordinate transformation, stitching, point cloud processing) and multispectral data (radiometric and geometric correction, mosaicking) was performed.
- Individual Tree Extraction: A multi-scale segmentation algorithm was used to precisely identify individual tree crowns and extract parameters like crown width and area.
- Spatial Connection: A crucial step involved precisely matching LiDAR-derived tree locations with multispectral segmentation objects. This was achieved using a BallTree nearest neighbor search algorithm with dynamic distance threshold filtering, ensuring accurate fusion of multi-source data.
- Biomass Model Construction: A Random Forest (RF) algorithm was employed to build a joint biomass model. This model incorporated tree height and crown width as key input features, leveraging the synergistic relationship between vertical growth and photosynthetic area.
- Model Validation: Spatial cross-validation (7:3 split and 5-fold) was used to ensure spatial independence and robust evaluation of the model's performance, mitigating issues of spatial autocorrelation.
Datasets
- UAV LiDAR Data: Collected using a D-LiDAR2000 payload, providing 3D structural information.
- UAV Multispectral Data: Collected using a D-MSPC2000 payload, providing spectral reflectance data.
- Ground Truth Data: Manual measurements of Diameter at Breast Height (DBH) were collected for validation.
- Sample Library: 70 sample plots were selected based on climatic zoning, vegetation types, and topography, forming a representative sample library for model training and optimization. This library is intended to be a replicable framework for similar regions.
Key Insights and Ideas
- Synergistic Power of Multi-Source Data: Combining LiDAR's 3D structural information with multispectral data's spectral characteristics is essential for accurate biomass estimation, overcoming the limitations of single-source data.
- Importance of Spatial Connection: Precisely linking LiDAR and multispectral data is critical. The BallTree algorithm with dynamic thresholding proved effective in achieving this, ensuring reliable data fusion.
- Random Forest for Complex Relationships: The RF algorithm excels at capturing the nonlinear relationships between vegetation structure, environmental factors, and biomass, making it well-suited for this application.
- Joint Model Superiority: A joint model using both tree height and crown width significantly outperforms single-factor models, highlighting the importance of considering multiple biophysical parameters.
- Climatic Zoning for Robustness: Using climatic zoning to delineate subzones and guide sample selection enhances model adaptability and generalization by accounting for regional variations.
Key Results
- The joint Random Forest model incorporating tree height and crown width achieved high R² values ranging from 0.839 to 0.934.
- This joint model significantly outperformed single-factor models, which had R² values of 0.156–0.314 (tree height) and 0.632–0.823 (crown width).
- The highest accuracy (R² = 0.934) was observed in the Hengshan–Wutaishan forest area.
- Accuracy was generally higher in forest-dominated zones and lower in areas with significant human disturbance or complex vegetation types.
- The spatial connection algorithm achieved an average matching rate of 73.98–82.01%.
- The model demonstrated good unbiasedness, with mean residuals close to zero across most vegetation zones.
Limitations and Future Work
- Spatial Matching Errors: Long-tail distributions of spatial matching errors were observed in specific terrain areas, particularly high-density building zones. Future research should focus on more precise matching methods for these areas.
- Species-Specific Differences: The current generalized model may exhibit prediction biases for single-species scenarios, especially in areas with complex species composition. Future work should aim to develop models that encompass diverse ecosystems (herbaceous, shrub, and tree) and integrate hyperspectral data for more detailed species classification.
- Influence of Human Disturbance: Vegetation fragmentation due to human activities can degrade model performance. Incorporating non-vegetation parameters like land-use classification and human disturbance intensity is suggested.
- Computational Constraints: Variables like spectral characteristics, topographic factors, and annual precipitation were not directly included due to computational constraints. Future research could explore their inclusion.
- Dynamic Monitoring: Expanding the spatial and temporal dimensions of the sample database is recommended for dynamic monitoring and spatiotemporal integrated models.
Summary updated