Classification of vegetation types in saline–alkali grassland of the Songnen Plain using multispectral UAV imagery and ensemble learning

Accurate classification of saline-alkali grassland vegetation presents considerable challenges due to high inter-class spectral similarity and severe class imbalance, which significantly limit traditional remote sensing approaches. This study introduces a stacking ensemble learning framework integrating high-resolution multispectral unmanned aerial vehicle (UAV) imagery for grassland vegetation classification in the Songnen Plain. Field surveys and UAV data were collected from Anda and Fuyu regions in Heilongjiang Province, China, focusing on three dominant vegetation types: Leymus chinensis (LC), Calamagrostis epigejos (CE), and Scirpus triqueter (ST). The framework combines Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP) with Logistic Regression (LR) as a meta-learner. The approach utilizes 192-dimensional feature vectors encompassing multiscale spectral indices and textural descriptors. To address class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) and data augmentation techniques were implemented. Five-fold cross-validation validated the framework’s robustness, yielding a mean macro-average F1 score (Macro-F1) of 85.79% (standard deviation, StD = 1.60%), demonstrating highly consistent performance across data partitions. Feature importance analysis revealed the mean Red Edge Chlorophyll Index (ReCI) calculated within 5 × 5 (L) and 7 × 7 (XL) windows as the most discriminative variables, with F-statistics exceeding 110, confirming physiological relevance for differentiating salt-tolerant vegetation types. Spectral features dominated classification performance, contributing 72.10% compared to 27.90% from textural features. This framework demonstrates the potential of ensemble learning approaches for saline-alkali grassland vegetation classification, though further development is needed for robust operational deployment. • A stacking ensemble framework classifies saline-alkali vegetation with 85.79% Macro-F1. • Red-edge spectral indices (ReCI) at L and XL scales are the most decisive features (F-statistic > 110). • Spectral features (72.10%) are more critical for classification than texture (27.90%). • Data augmentation is indispensable, boosting the Macro-F1 score by over 10.00%. • The framework outperforms deep learning (e.g., Transformer 1D, 1D-ResNet) on this small dataset.