Soil salt contents highly affect the ecology and human health. However, due to the complexity of soil, the evaluation has been reliant on analytical testing laboratory analysis. Such a practice requires the operation of high-cost equipment and long analysis time. Thus, the persistent monitoring for the soil quality is difficult.

This study investigated a rapid salt contents determination method through the combination of Ultraviolet-Visible (UV-Vis) Spectroscopy and Electrochemical Impedance Spectroscopy (EIS), and the application of convolutional neural network (CNN). For the spectroscopic measurement, 143 aqueous salt samples were prepared with various salts including Ca²⁺, K⁺, Na⁺, Cl⁻, Br⁻, SO₄²⁻, and HCO₃⁻, and different concentrations thereof. The spectral data obtained from quadruple measurement on UV-Vis spectroscopy and EIS were individually analyzed to evaluate the candidacies of them. Also, through the principal component analysis (PCA), the effect of fused data incorporating UV-Vis and EIS spectra was qualitatively and quantitatively assessed.

A CNN network was designed to predict seven ionic contents at once such as Ca²⁺, K⁺, Na⁺, Cl⁻, Br⁻, SO₄²⁻, and HCO₃⁻. The network was tested with individual and fused spectral datasets. The evaluations of the network performance were done with the test sets separated prior to the training process. The prediction tests showed the feasibility of UV-Vis and EIS spectroscopies in predicting ionic contents with RMSEP of 112.58 and 81.83 mmol/kg each. However, the prediction performance with the fused spectra was significantly improved with 48.58 mmol/kg RMSE. Also, the CNN was compared with other machine learning models, PLSR and random forest. PLSR and random forest resulted in 96.47 and 95.47 mmol/kg RMSEP each. From the comparison, the prediction capability of the proposed CNN network became further pronounced with 48.58 mmol/kg RMSEP.

The detection of various ionic contents enables detailed information about the subjected area, providing chances for the development of ion-specific remediation methods thereof. The required spectroscopy devices are portable and rapid and can be operated on-site. This will facilitate a more frequent monitor of soil conditions in timely manner. Moreover, more datasets and the optimization of the CNN network will further improve the performance.