Stacking ensemble learning coupled with multi-source remote sensing data: Enhancing soil salinity inversion accuracy in barley-cultivated salinized soils

Soil salinization poses an increasing threat to arable land quality, crop growth, and regional biodiversity; therefore, accurate and efficient acquisition of salinity information is essential for remediation. Previous studies have primarily focused on single data sources or algorithms, whereas the potential of integrating multi-source remote sensing data with multiple machine learning models remains largely underexplored. In this study, quantitative experiments were conducted across gradients of soil salinity under barley cultivation to acquire visible light (RGB), multispectral (MS), and thermal infrared (TIR) imagery, together with measurements of apparent electrical conductivity of the soil profile. A total of 144 feature variables, including band brightness, spectral reflectance, and derived color indices, temperature indices, and texture features, were extracted from remote sensing image, screened and optimized using coordinated approach integrating random forest importance and recursive feature elimination (RF-RFE), and incorporated into seven data fusion schemes. Four machine learning models and three ensemble learning strategies (Average, Weighted, and stacking of two layers with ridge regression as the meta-model (St-RR)) was systematically established and assessed using cross validation with five folds and multiple evaluation metrics. Results showed that extreme learning machine achieved the highest accuracy for surface soil (R²=0.75, RMSE=0.69 %, MAE=0.45, RPIQ=0.76, RPD=1.98), whereas gaussian process regression performed best for root-zone soil (R²=0.73, RMSE=1.06 %, MAE=0.68, RPIQ=1.08, RPD=1.93). All three ensemble learning strategies improved estimation accuracy compared with single models, with St-RR achieving the most notable enhancements (surface soil: R² increases of 6.0 %–7.6 % and RMSE reductions of 5.1 %–6.3 %; root-zone soil: R² increases of 6.1 %–9.4 % and RMSE reductions of 8.3 %–10.6 %). Among all data fusion schemes, the St-RR based on the RGB+MS fusion achieved the highest accuracy for surface soil (R² = 0.76, RMSE = 0.67 %, MAE = 0.43, RPIQ = 0.79, RPD = 2.04) and root-zone soils (R² = 0.77, RMSE = 0.99 %, MAE = 0.65, RPIQ = 1.16, RPD = 2.08). Overall, these findings demonstrate the effectiveness of multi-source remote sensing fusion combined with stacking ensemble learning for accurate soil salinity estimation, providing robust technical support for the management of soil salinization at a fine scale. • Fusion of multi-source remote sensing data improves soil salinity estimation accuracy. • Random forest combined with recursive feature elimination effectively selects features relevant to soil salinity. • Ensemble learning outperforms single machine learning models in salinity estimation. • The study provides a robust, fine-scale approach for monitoring salinization in croplands.