An ensemble learning method for soil nutrient prediction based on multi-source data fusion

In response to the issue of limited predictive accuracy in digital soil mapping due to insufficient model matching, this paper proposes a soil attribute prediction method based on multi-source data fusion, named MSD-Stacking. This method innovatively constructs a Stacking ensemble strategy that integrates spectral data and environmental covariates, utilizing their complementarity to build diverse base learners and overcoming the limitations of a single data source. Moreover, it employs a performance weighting mechanism to enhance the weight of well-performing base learners in the ensemble, thus optimizing the integration effect. The algorithm demonstrated in soil pH prediction experiments on the LUCAS 2009 dataset shows that, compared to the Stacking model and traditional machine learning algorithm models, MSD-Stacking achieves an average increase of 42.38% in R2, a reduction of 31.74% in RMSE, and a reduction of 33.26% in MAE. This method provides a more precise and stable technical solution for digital soil mapping, advancing soil attribute prediction towards efficiency and intelligence.