Enhancing soil geographic recognition through LIBS technology: integrating the joint skewness algorithm with back-propagation neural networks

Abstract

The meticulous task of soil region classification is fundamental to the effective management of soil resources and the development of accurate soil classification systems. These systems are crucial for the targeted restoration, safeguarding, and enhancement of land resources. In this research, we introduce an innovative soil classification model that combines the Joint Skewness (JS) algorithm, which is grounded in tensor theory, with a Back-Propagation Neural Network (BPNN). This combination is utilized for the rapid categorization of soil samples in specified areas, making use of spectral data from Laser-Induced Breakdown Spectroscopy (LIBS). The process begins with the application of JS to identify key variables, followed by the optimization of the JS-BPNN model's structure. The effectiveness of the model is then evaluated using metrics such as the confusion matrix, Kappa coefficient, and precision, which all highlight the model's reliability. Our experimental results validate the use of JS in filtering LIBS spectral features, effectively minimizing unnecessary data while preserving the spectral data's intrinsic physical characteristics. This leads to a significant enhancement in the model's analytical capabilities. The JS-BPNN model has demonstrated remarkable classification accuracy, achieving a peak accuracy of 99.8% on the test dataset. To further validate the JS approach for reducing data dimensionality and emphasize the superiority of the JS-BPNN model, we conducted a comparative analysis with other algorithms, such as k-Nearest Neighbor (KNN), Random Forest (RF), and Support Vector Machine (SVM), for the classification and recognition of soil geographic regions. The results confirm that the JS algorithm is a potent method for reducing the dimensionality of LIBS spectral data, and for different classification models, there are different optimal characteristic variables, with the JS-BPNN model proving to be exceptionally effective in soil classification and recognition tasks.