Supporting soil and land assessment with machine learning models using the Vis-NIR spectral response

Soil Vis-NIR spectral response had been widely proposed as an alternative to costly and time-consuming laboratory determination of soil physical and chemical properties. However its use for measuring soil quality index directly has not been well explored. This study compares the effectiveness of different machine learning models on a large spectral library using a database collected by the European Union project “Land Use and Coverage Area frame Survey” (LUCAS). Three approaches to predicting mineral soil features by processing their spectral response for the Vis-NIR range were tested. Prediction models of clay content, pH in CaCl2, organic carbon (SOC), calcium carbonate (CaCO3), nitrogen (N), and cation exchange capacity (CEC) were analyzed. Three types of models were assessed: a Stacked AutoEncoder, a convolutional neural network, and a stack model composed of a set of multilayer perceptron algorithms with two different regression estimation solutions. Modeling with CNN was identified as the optimal solution. Similar, and in some cases, better results can be obtained from ensembles of machine learning algorithms. The estimates of soil characteristics made with the help of the Stacked AutoEncoder showed the greatest errors. The use of soil feature estimates to support soil and land classification was also analyzed. An indicator describing the state of the topsoil is presented, which assists the objective classification of soils. The research showed that the accuracy of the estimation of the proposed Topsoil Quality Index (TQI) estimated directly based on Vis-NIR spectral response and indirectly based on estimated values of selected soil features is practically identical. The research confirms the suitability of Vis-NIR spectroscopy for topsoil assessment.