Rapid and accurate detection of total nitrogen in the different types for soil using laser-induced breakdown spectroscopy combined with transfer learning

Abstract

Precision fertilizing is crucial not only for enhancing fertilizer efficiency but also for protecting the environment. The rapid sensing of total soil nitrogen (TN) constitutes a key aspect of precision fertilization. Currently common methods, such as the Kjeldahl method, are not suitable for on-site applications. Laser-induced breakdown spectroscopy (LIBS), celebrated for its expeditious data acquisition and high precision, has seen widespread deployment in rapid soil sensing. However, the time-consuming sample preprocessing stage restricts the on-site application of LIBS. In this study, we employed a powder adhesion (PA) method to shorten the preprocessing cycle to 3 min. A transfer learning approach named TransLIBS is introduced to ensure the estimation performance of PA. Compared to the calibration model directly developed on the target domain, the transferred model by TransLIBS elevates $R_V^2$ by 0.134 and diminishes ${\mathrm{RMSE}}_V$ by 0.312 g kg⁻¹. The F-test method is leveraged to identify active variables, and feature map visualization is employed to interpret the transfer mechanism of the TransLIBS approach. The visualization results highlight the most influential variables situated in the 212–310 nm and 391–395 nm range. Transfer learning has advanced the application of LIBS in soil, providing more opportunities for on-site LIBS detection.