A Super Learner Ensemble to Map Potassium Fixation in California Vineyard Soils

Abstract

Potassium (K) deficiency in wine grapes results in reduced vine growth, premature leaf drop, and yield and color loss. K can be fixed in the interlayer of certain clay minerals in a process called K fixation. This, coupled with management and other soil factors, leads to high spatial variability in soil K. In the Lodi American Viticulture Area (AVA) management of winegrapes is complicated by a mix of K fixing soils and non-K fixing soils. Identification of the spatial distribution of K fixation and availability can help disentangle the complexity of K management in the region. Soil samples (n=107) were collected, analyzed for K fixation, K availability and cation exchange capacity (CEC), and aggregated into two depths (0-30 cm and 30-100 cm). Soil samples were intersected with remotely sensed proxies for the soil forming factors and existing soil survey data and used as training data for a “super learner” ensemble or combination of base models, including random forest (RF), extreme gradient boosting (XGB) and Cubist. Base models were combined via model averaging (each model weighted by its R2) or model stacking (linear combination of base models via OLS regression), and model performance was compared. We report generating mapped uncertainties from a super learning framework by generating bootstrapped realizations of each base model and weighting each bootstrapped base model map via the β-coefficients generated in the ensemble fitting step.  This generated bootstrapped maps of the super learner, which were utilized to generate upper and lower 90% prediction limits.  For K fixation in the 0-30 cm depth, RF outperformed other models (R2=0.42), whereas a linear combination of all base models performed best in the in 30-100 cm depth (R2=0.41). Results improved for K availability with an R2 of 0.48 in the 0-30 cm horizons and R2 of 0.46 in the 30-100 cm horizons. Finally, predictions for CEC were superior, with R2 of 0.71 for the surface, and 0.51 for the 30-100 cm depth. Results suggest that K fixation and availability can be predictively mapped with marginal success, while CEC is amenable to a DSM framework.  CEC is tied more to soil genesis and formation, while K fixation and availability are affected by K fertilization. Finally, we compared the DSM K fixation map to an existing soil landscape model (SLM) K fixation map, to facilitate discussion of the connection between pedogenic state factors, soil forming processes and morphologies in soil mapping.

Keywords: Potassium fixation, K fixation, vineyard soils, super learner, ensemble mapping, digital soil mapping, pedogenesis, pedology, uncertainties, soil landscape model