Fusion of visible near‐infrared and mid‐infrared data for modelling key soil‐forming processes in loess soils

Abstract

AbstractPedogenic weathering of aeolian materials plays an essential role in global nutrient and carbon cycling. A quick and effective approach for temporally high‐resolution and spatially large‐scale pedogenic investigations has long been needed. Here, we used visible and near‐infrared, mid‐infrared and sensor‐fused data to predict pedogenesis‐related soil properties (i.e., grain‐size distribution, clay‐mineral properties and geochemical ratios) in a thick loess sequence in central China. Sensor fusion was achieved at three different levels: (1) direct combination of spectral parameters (low‐level fusion; Fusionpara); (2) combination of spectral features selected by principal component analysis (middle‐level fusion; FusionPCA) and (3) fusion using outer product analysis (high‐level fusion; FusionOPA). Sensor‐fusion generally improves the model predictions for all soil properties, with increases in the values of the model efficiency coefficient (MEC) of 8% and the performance‐to‐interquartile range of 12%. Whole‐soil properties are optimally predicted with the Fusionpara dataset using the random forest algorithm, yielding a mean MEC of 0.85. Spectral parameters D2200/D1900 and AS2200 are demonstrated as promising new pedogenic proxies, with higher values indicating more intense hydrolysis during pedogenesis. Spectral proxies along the loess sequence suggest that intense soil formation occurred during warm and humid interglacial periods when the East Asian summer monsoon intensified. The sensor‐fusion technique improved model performance for assessing mineral transformations and chemical weathering intensity, providing a quick and efficient means of interpreting pedogenic evolution, especially in the case of spatially large‐scale soil investigations.Highlights Fusion of VNIR and MIR spectra to model loess pedogenesis. Spectral modeling provides a robust surrogate for pedogenic evolution studies. Sensor fusion improves the performance of chemometric models. Sensor‐fusion models can reconstruct loess pedogenesis.