Mixed-Effects Modeling for Soil Property Prediction Using Chemometric and Environmental Data.

Traditional chemometric models use solely spectral data to predict various soil properties. We explored a new mixed methods methodology from the genomics field, namely Hierarchical Generalized Linear Model (HGLM), incorporating both fixed effects, i.e. environmental covariates, and random effects, i.e., site-specific spectral data, into the modeling process. It is unknown how such a mixed method approach performs in soilscapes with diverse environmental and spectral characteristics.

This study used data acquired from two agricultural villages in southern India (separated by 300 km). Over 250 soil samples were collected from each village, which were measured for texture, soil carbon, and available nutrients. The soil samples were also scanned in the mid-infrared (MIR) range (2,500 – 25,000 nm). Spatial environmental datasets were also collected that highlight topography, phenology, and management. The objectives of the study were to (1) predict soil properties with the HGLM for each village using the MIR data as random effects and spatial data as fixed effects, (2) predict soil properties using the pooled village datasets, and (3) compare the results with a standard chemometric method.

Preliminary results show that incorporating fixed effects improve prediction results over a standard chemometric method. The HGLM method may provide a means of overcoming the issue of combining environmental (spatial) covariates with high-dimensional, multi-collinear spectral (point) data within a model rather than using the environmental covariates to cluster or stratify data prior to modeling. This method may also be useful for modeling environmental covariates with other high-dimensional spectroscopic datasets. This research highlights the importance of exploring new methods from other fields, especially those that also employ statistical modeling, that can handle the diverse sets of data collected to study soils.