Dimensionality reduction of soil vis–NIR spectra: implications for soil health assessment and mapping

The measured soil spectrum in the visible and near-infrared (vis–NIR) range contains information on various soil physical properties, as well as mineralogical and chemical composition. Therefore, this approach can serve as a rapid and cost-effective alternative for assessing soil health. However, the raw soil spectrum is high-dimensional and less suitable for modeling purposes. In this study, we compared the performance of four methods, including PCA, kernel PCA (KPCA), autoencoders (AE), and convolutional autoencoders (CNN-AE), based on their reconstruction error, which directly measures information loss during dimensionality reduction. We used a comprehensive soil dataset from Denmark comprising 7,009 vis–NIR spectra covering a wide range of land uses and soil types. Our analysis shows that reconstruction error decreases as the number of latent variables (LVs) increases, as expected, due to the greater capacity to preserve information. Notably, across all latent dimensionalities, nonlinear methods consistently outperformed PCA. At low latent dimensionalities (5–10 LVs), nonlinear methods achieved approximately 45–55% lower reconstruction error than PCA. As the number of LVs increases, this performance gap decreases. However, PCA still consistently shows lower performance than the other methods. Given the strong representation of soil chemical, physical, and hydrological properties in the spectra, we mapped the latent variables across Denmark using diverse spatial predictors. The mapped latent variables captured complex, nonlinear soil–landscape relationships that reflect key mineral and organic soil components. Overall, our results suggest that spectral representations could serve as a scalable approach to support national soil mapping.