Spectral-guided ensemble modelling for soil spectroscopic prediction

Ensemble modelling (EM) has been increasingly used in soil information prediction by spectroscopic techniques to enhance model robustness and improve model performance. This approach is usually implemented by fitting a new model using the predictions from several predictive models, and then outputting new predictions. Since the prediction error associated with each model are randomly distributed, the useful information derived from the predictions of each predictive model is somewhat limited. In this study, we proposed a new approach, namely spectral-guided ensemble modelling (S-GEM), to improve soil spectroscopic prediction by including spectral information in EM. Taking LUCAS Soil 2009 data as an example, our results showed that S-GEM performed better than EM using Granger-Ramanathan (a gain of R2 of 0.04–0.05) as well as the best classic model including partial least squares regression, Cubist and random forest (a gain of R2 of 0.08–0.09) for predicting soil organic carbon, clay and pH using vis-NIR spectra. Therefore, we suggest that S-GEM has a high potential to improve soil spectroscopic prediction over the conventional EM, and therefore provides more accurate soil information for monitoring soil status and changes over space and time using digital soil mapping. In addition, the idea of including auxiliary information in EM can also be extended outside of pedometrical applications for improving predictive ability.