Applying convolutional neural networks (CNN) for end-to-end soil analysis based on laser-induced breakdown spectroscopy (LIBS) with less spectral preprocessing

Abstract

The high variation of raw laser-induced breakdown spectroscopy (LIBS) caused by soil heterogeneity seriously reduces the accuracy and stability of the spectral analysis. Therefore, the conventional chemometrics for spectral analysis requires seeking an appropriate spectral preprocessing by a trial-and-error method before modeling, which resulted in a mutable performance. To settle this problem, the convolutional neural network (CNN), a type of deep learning approach with the advantage of end-to-end, was applied to predict soil type and soil properties based on the non-preprocessed LIBS spectra. The results indicated, when compared to conventional partial least squares (PLS), that the CNN models presented equal classification accuracy but they decreased the root mean square error in the validation set (RMSEV) by 1.48%, 4.97%, 9.56%, 10.05%, and 2.90% for pH, soil organic matter (SOM), total nitrogen (TN), total phosphorus (TP), total potassium (TK), respectively. In addition, the CNN models performed better in preventing overfitting than the conventional PLS combined with various spectral preprocessing approaches. The multi-task of CNN models also further improved the prediction of TN due to its capacity to learn inherent structures from spectra. The sensitivity analysis of spectral variables revealed that the CNN model with the Inception module discovered both the local and high abstracted features compared with other CNN models. In conclusion, the CNN architectures showed potential to end-to-end deal with raw soil LIBS spectra.