Rapid estimation of soil Mn content by machine learning and soil spectra in large-scale

Abstract

Manganese (Mn) is an essential element in both plants and the human body; however, traditional methods for monitoring Mn in soil are costly and inefficient. As such, it is necessary to establish a model for environmental research uses that can accurately predict soil Mn content over large areas. This study aims to develop a multilayer perceptron (MLP) model capable of accurately predicting Mn content in diverse soils based on visible and near-infrared (VNIR) spectroscopy. A dataset containing 18,675 soil samples compiled from the Land Use and Coverage Area Frame Survey was used to train and adjust the model, following which the optimal model was applied globally. The correlation coefficient, root mean square error, and mean absolute error values for the optimal model on the test set were 0.76, 140.52, and 97.30, respectively. Feature importance analysis revealed crucial spectral bands at approximately 1400, 2200, 2300, and 2400–2500 nm, which enabled the Mn content to be estimated. The presence of these spectral bands indicates that clay minerals, H2O, and OH− groups had significant influence on Mn content. The MLP model developed can effectively identify regions with potentially high or low soil Mn content. With improvement of the spectral database, this model can provide effective assistance in evaluating soil Mn distribution at the global scale.