Assessing the Information Potential of MIR Spectral Signatures for Prediction of Multiple Soil Properties Based on Data from the AfSIS Phase I Project

The aim of the study was to assess the predictive potential of mid-infrared (MIR) spectral response in the estimation of 60 soil properties. It is important to know the accuracy limitations in estimating various soil characteristics using various models in conditions of high spatial variability of the environment. To fully assess this potential, three types of algorithms were used in modeling, i.e., partial least squares (PLSR), one-dimensional convolutional neural network (1DCNN), and generalized regression neural network (GRNN). The research used data from 19 sub-Saharan African countries collected as part of the Africa Soil Information Service (AfSIS) Phase I project. The repositories provide 18,250 MIR reflectance recordings and nearly two thousand analytical data records from the determination of many soil properties by reference methods. The modeled subset of these properties included texture (three variables), bulk density, moisture content at soil water characteristic curves (SWCC, 4 variables), total and organic C and total N content (3 variables), total elemental content (32 variables), elemental content in bioavailable forms (12 variables), electrical conductivity, exchangeable acidity, exchangeable bases, pH, and phosphorus sorption index. It is not possible to indicate a universal optimal prediction model for all soil variables. The best prediction results are provided by all regression models for total and organic C, total Fe, total Al and bioavailable Al content, and pH. For bulk density, total N and total K content satisfactory results are provided by specific model type. Many other properties, i.e., texture, SWCC, total Ga, Rb, Na, Ca, Cu, Pb, Hg content, and bioavailable Ca and K content, can be predicted with accuracies sufficient for some less demanding tasks.