Improved Grass Species Mapping in High-Diversity Wetland by Combining UAV-Based Spectral, Textural, Geometric Measurements

Accurate mapping of grass species in biodiverse ecosystems, such as wetlands, is critical for ecological protection. Rapid advancements in remote sensing have established satellite data as a critical tool for wetland grass species mapping; however, its relatively coarse spatial resolution and susceptibility to cloud contamination limit the distinction of co-occurring species at fine scales. While Unmanned Aerial Vehicle (UAV) remote sensing offers high resolution and operational flexibility, relying on single-source features is often insufficient for fine-scale wetland species mapping due to the spectral similarity of co-occurring species. On the other hand, the fusion of multi-source remote sensing features (i.e., spectral, textural, and geometric features) likely provides a promising solution for achieving accurate, fine-scale grass species mapping in biodiverse ecosystems. In this study, we developed a wetland grass species mapping framework integrating spectral, textural, and geometric features derived from UAV RGB and multispectral imagery. Using a dataset of 95,880 image objects representing 24 wetland grass species classes collected in two years in Dajiu Lake National Wetland Park of China, we evaluated three machine learning algorithms—Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost)—across various feature combinations. We found that while spectral features (i.e., red edge, normalized green–red difference index [NGRDI], and normalized difference vegetation index [NDVI]) (related to leaf pigment concentrations and cellular structures) exhibited the highest importance in wetland grass species mapping, textural (i.e., contrast) and geometric features (i.e., aspect ratio) significantly enhanced classification performance as complementary information, yielding improvements of up to 10.5% in overall accuracy (OA) and 0.103 in Macro-F1 scores. Specifically, the fusion of spectral, textural, and geometric features achieved optimal performance with an OA of 81.9% and a Macro-F1 of 0.807. Furthermore, the XGBoost model outperformed SVM and RF, improving OA by 9.4% and 2.8%, and Macro-F1 by 0.08 and 0.035, respectively. By identifying the optimal feature combination and machine learning algorithm, this study establishes an accurate method for wetland grass species mapping, offering new opportunities for ecological assessment and precision conservation in biodiverse landscapes.