New UAV-Based Framework Achieves 92.77% Accuracy in Karst Wetland Vegetation Classification

A new study published in the Journal of Remote Sensing demonstrates a significant advancement in wetland ecosystem monitoring through an adaptive ensemble learning framework that combines hyperspectral and LiDAR data collected by unmanned aerial vehicles (UAVs). The research, conducted in China's Huixian Karst Wetland, achieved up to 92.77% accuracy in vegetation species classification, substantially outperforming traditional remote sensing approaches and offering new capabilities for biodiversity conservation and carbon cycle monitoring.

Karst wetlands represent globally significant ecosystems that regulate water resources, store carbon, and support rich biodiversity. However, accurate vegetation mapping in these environments has been challenging due to intricate species composition and similar canopy spectra among different plants. Traditional field surveys are costly and spatially limited, while conventional remote sensing methods lack the resolution needed for species-level classification. The integration of complementary optical and structural data has emerged as a necessary approach for precise vegetation mapping in these complex environments.

Researchers from Guilin University of Technology and collaborators developed an adaptive ensemble learning stacking (AEL-Stacking) framework that merges hyperspectral imagery with LiDAR point-cloud data. The study, published with DOI 10.34133/remotesensing.0452, demonstrates how this integrated approach achieves superior classification accuracy while providing interpretability through local interpretable model-agnostic explanations (LIME). The framework combines Random Forest, LightGBM, and CatBoost classifiers within a grid-search-optimized adaptive system that uses 70% of data for training and 30% for testing, supported by 10-fold cross-validation.

The research team conducted field surveys in the Huixian Karst Wetland, where UAV flights equipped with Headwall Nano-Hyperspec and DJI Zenmuse L1 LiDAR sensors collected over 4,500 hyperspectral images and dense point clouds at 208 points per square meter. The integrated dataset covered 13 vegetation types including lotus, miscanthus, and camphor trees. Through recursive feature elimination and correlation analysis, researchers selected 40 optimal features from more than 600 variables, with LiDAR-derived digital surface model (DSM) variables proving particularly valuable for distinguishing species with distinct vertical structures.

Results showed that combining hyperspectral and LiDAR data achieved overall accuracy between 87.91% and 92.77%, surpassing single-data approaches by up to 9.5%. The AEL-Stacking model outperformed both conventional ensemble methods and deep-learning algorithms by 0.96% to 7.58%. Hyperspectral vegetation indices such as NDVI and blue-edge parameters enhanced recognition of herbaceous species, while LIME analysis revealed DSM and blue spectral bands as the most influential features. Lotus and miscanthus achieved classification F1-scores above 0.9, with the model significantly reducing misclassification between morphologically similar species.

"Our approach bridges the gap between spectral and structural sensing," said Dr. Bolin Fu, corresponding author of the study. "By combining UAV hyperspectral and LiDAR data through adaptive ensemble learning, we achieved both precision and interpretability in vegetation mapping. The framework not only improves species recognition in complex karst environments but also provides a generalizable tool for ecological monitoring and habitat restoration worldwide."

The research demonstrates a scalable and explainable approach for high-resolution wetland mapping that could potentially be applied to forest, grassland, and coastal ecosystems. Future work will focus on integrating multi-temporal UAV observations and satellite data fusion to monitor seasonal vegetation dynamics and climate-driven changes in wetland health. By enhancing the transparency and accuracy of AI-driven ecological models, this research supports global biodiversity conservation efforts and carbon neutrality initiatives while providing detailed vegetation maps essential for ecosystem monitoring and restoration planning.