Spatial variability of soil salinity in coastal saline-alkali farmlands: A novel approach integrating a stacked model with the reconstructed in-situ hyperspectral feature

• An adaptive stacker model was introduced to reconstruct in-situ hyperspectral features. • Reconstructed hyperspectral features effectively capture spatial variability of soil salinity . • Soil salinity exhibits a gradient distribution, increasing from inland areas towards the coast. • Climate and vegetation factors synergistically exert a profound influence on soil salinity. Dynamic and rapid hyperspectral monitoring of soil salinity is of great significance for the sustainable agriculture 4.0 development. However, it is difficult to retrieve the spatial variability of regional soil salinity by in-situ hyperspectral (ISH) data due to the defects of non-spatial continuity. For the end, we developed a novel ISH reconstruction method that integrates satellite-borne hyperspectral (SBH) with a trio of model learning (ML) (i.e. gradient boosting decision tree (GBDT), random forest (RF), and deep forest (DF)) into a partial least square (PLS) stacker to reconstruct the feature bands of ISH. This method enables precise predictions of the spatial variability in soil salinity across coastal saline agricultural lands. The stacker achieves a fitting accuracy (R 2 ) of over 0.8 between SBH and ISH features in comparison with single machine learning models. Compared with ISH (R 2 = 0.82 and RMSE = 1.7 g kg −1 ) and SBH features (R 2 = 0.39 and RMSE = 3.7 g kg −1 ), the tree-structured Parzen estimator-extreme gradient boosting (TPE-XGB) model based on reconstructed in-situ hyperspectral (RISH) features achieves the highest prediction accuracy (R 2 = 0.86 and RMSE = 1.0 g kg −1 ) for soil salinity. The hyperparameter combination of the XGB model optimized by the TPE algorithm exhibits strong heterogeneity and high predictive performance. SHapley additive interpretation (SHAP) analysis results reveal that visible bands contribute most significantly to prediction accuracy. The spatial distribution of soil salinity concentrations in the coastal saline areas exhibits a gradient trend that decreases from the coast to the inland. Moreover, climate and vegetation factors dominate the spatial variability of soil salinity, exhibiting a synergistic effect as 85.26 % based on the variogram partitioning (VP) and hierarchical partitioning (HP) method. In summary, the stacker-assisted ISH feature reconstruction method is highly feasible for predicting spatial variability of soil salinity, providing technical support for precise monitoring and intelligent management of sustainable agriculture 4.0.