High-resolution mapping of compressed vis-NIR spectral data of Danish topsoils

Visible and near-infrared (vis-NIR) spectroscopy enables rapid and cost-effective soil characterization, supporting the estimation of soil chemical, physical, and biological properties. Most applications, however, focus on using spectroscopy to predict individual soil properties and augment conventional laboratory analyses, while the potential of spectral information to study soil variation in space remains largely unexplored. Latent variables derived from the dimensionality reduction of spectral data offer a compact representation of soil variability, capturing multiple soil properties simultaneously. When spatially predicted, these latent variables can provide a means to investigate soil-landscape relationships in a spatially explicit context.
This study aimed to (i) model the spatial variation of latent variables derived from compressed topsoil vis-NIR spectral data across Denmark at 10 m resolution using a digital soil mapping approach; (ii) identify the drivers, in relation to SCORPAN factors, controlling their spatial variation; and (iii) examine how the predicted latent variables represent the soil-landscape relationships across Denmark. An earlier study compressed the data using principal component analysis, kernel principal component analysis, shallow autoencoders, and convolutional autoencoders. To predict the latent variables in space, we used 32 predictors comprising harmonized environmental layers representing climate, relief, and parent material. A weighted machine learning algorithm was used to model each latent variable, with bootstrap resampling providing pixel-level uncertainty estimates. Overall, climate was the main driver across all methods, followed by topography and parent material, mainly the extent of clay till deposits.
By integrating compressed spectral data with diverse spatial predictors, we captured complex, non-linear soil-landscape relationships. The resulting high-resolution maps offer a spatial description of soil variation in relation to soil-forming factors and spectral signature. Our maps can be used to improve point-based predictions of soil properties or provide enhanced covariates to support digital soil mapping of soil properties and classes.