A hybrid machine learning approach for classifying spectrally similar soils using laser-induced breakdown spectroscopy

Abstract Accurate classification of geographic regions of soil is imperative for precision agriculture, targeted land management, and other related applications. Often, soil samples for such analyses exhibit high similarity in features, especially from geographically neighboring regions, which poses a challenge for accurate classification of soil geographic regions based on laser-induced breakdown spectroscopy (LIBS) technology. To address this issue, we propose a hybrid classification framework that integrates variable importance measure-random forest (VIM-RF) with a back propagation neural network (BPNN). The LIBS spectral data used in this study were collected from soil samples with close geographical origins and high inherent similarities. The VIM-RF algorithm was employed to identify and retain the most discriminative spectral variables, which are essential for distinguishing these similar soil classes. Based on the value of VIM, 30 spectral variables were retained, which covered these trace elements, such as Fe, Mn, Ti, as well as major elements like K and N in the soil. Subsequently, the screened variables were used to train the BPNN model for classification, and after 10-fold cross-validation optimization, achieved a classification accuracy of 99.6%±1.21%. The experimental results demonstrate that the proposed VIM-RF-BPNN framework has superior classification performance, and the effectiveness of the VIM-RF method in screening highly relevant features from complex and highly similar LIBS data. When combined with BPNN, it can significantly improve the classification accuracy. This study highlights the potential of VIM-RF as a powerful feature selection strategy and confirms the effectiveness of the VIM-RF-BPNN approach for accurate soil classification based on LIBS, particularly in scenarios involving spectrally analogous or overlapping soil profiles.