Improving the estimation of alpine grassland fractional vegetation cover using optimized algorithms and...

Area of Science:

• Remote Sensing
• Ecology
• Geospatial Analysis

Background:

• Fractional vegetation cover (FVC) is a key parameter for monitoring alpine grassland ecosystems on the Qinghai-Tibetan Plateau.
• Accurate FVC products are crucial for understanding ecosystem health and dynamics.
• Unmanned aerial vehicle (UAV) data combined with satellite imagery offers potential for high-precision FVC estimation.

Purpose of the Study:

• To develop and optimize algorithms for accurate FVC inversion in alpine grasslands.
• To evaluate the effectiveness of multi-dimensional feature sets versus traditional vegetation indices.
• To identify optimal feature selection and machine learning algorithms for FVC estimation.

Main Methods:

• Constructed a multi-dimensional feature set including spectral bands, vegetation indices, and topographical factors.
• Employed feature selection algorithms (Boruta, SFS, PI-RFE) to identify optimal subsets.
• Evaluated four machine learning algorithms (including Random Forest) for FVC inversion accuracy, sensitivity, and efficiency.
• Optimized hyperparameters for selected machine learning models.

Main Results:

• Random Forest (RF) algorithm showed best performance using typical vegetation indices (R²: 0.861).
• Multi-dimensional features significantly improved FVC inversion accuracy for all tested algorithms (RF R² increased to 0.890).
• Permutation Importance-Recursive Feature Elimination (PI-RFE) demonstrated superior dimensionality reduction.
• Optimized RF algorithm with feature selection achieved the highest accuracy (R²: 0.917, RMSE: 7.9%).

Conclusions:

• An optimized algorithm utilizing a multi-dimensional feature set provides highly precise FVC inversion.
• This approach is vital for quantitative ecological monitoring of alpine grasslands.
• The study offers a robust methodology for generating high-quality FVC products.