Abstract
Potassium is a critical macronutrient for plant growth, yet accurately and rapidly estimating its content in karst regions remains challenging due to complex terrestrial conditions. To address this, we collected leaf potassium content and reflectance data from 301 plant samples across nine karst regions in Guangxi Province. Our results showed that hybrid models combining Partial Least Squares Regression (PLSR) with three machine learning algorithms-Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP)-namely PLSR-RF, PLSR-XGBoost, and PLSR-MLP, demonstrated exceptional accuracy in estimating leaf potassium content. Validation coefficient of determination (R²) values reached 0.89, 0.94, and 0.96, respectively-representing improvements of 206%, 147%, and 108% over standalone algorithms. This performance gain was attributed to rigorous overfitting control: PLSR's dimensionality reduction synergized with ensemble machine learning (RF, XGBoost, MLP) to eliminate redundant spectral features while retaining predictive signals. Furthermore, fractional differentiation preprocessing significantly improved the correlation between spectral reflectance and potassium content, enhancing model robustness. Two spectral regions (700-1100 nm, 1400-1800 nm) were identified as key predictors, aligning with known potassium-related biochemical absorption features. Collectively, the integration of these strategies offers a robust framework for nutrient monitoring in ecologically fragile karst ecosystems.