Visible and near-infrared spectroscopy coupled with machine learning to quantify oxalate-extractable Al and Fe and identify andic soil properties for Japanese forest soils

The combined content of oxalate-extractable aluminum (Alo) and 1/2 oxalate-extractable iron (Feo), expressed as Alo +1/2Feo, is a key factor for identifying andic soil properties. Specifically, a value of ≥20 g kg−1 for this variable is needed for andic soil properties in soils with less glass content. In this study, we thus developed a model to estimate the Alo and Feo contents in soils using visible and near-infrared spectroscopy coupled with machine learning. We evaluated the performance of three algorithms, namely partial least square regression, random forest (RF), and convolutional neural network (CNN). The CNN model outperformed the other models in estimating the Alo (R2 = 0.921) and Feo (R2 = 0.723) contents, with mean absolute errors of 2.3 and 2.0 g kg−1 for Alo and Feo, respectively. In evaluating whether the Alo +1/2Feo content of 2245 soil samples was ≥20 g kg−1, the RF and CNN models produced only 124 and 88 misclassifications, respectively. Our method can serve as a time-efficient and chemical-free approach for identifying andic soil properties.