Laser-Induced Breakdown Spectroscopy (LIBS) for tropical soil fertility analysis

Highlights

• LIBS satisfactorily predicted eight out of nine fertility attributes assessed

• iSPA-PLS proved to be the best method for predicting fertility attributes with LIBS

• The multivariate models tested are less matrix-matching dependent

• Multivariate models using LIBS data allowed satisfactory predictions of ex-P

Abstract

The Laser Induced Breakdown Spectroscopy (LIBS) is a promising technique for soil fertility analysis in a rapid and environmentally friendly way. This application requires the selection of an optimal modelling procedure capable of handling the high spectral resolution of LIBS. This work aimed at comparing different modelling methods of LIBS data for the determination of key fertility attributes in Brazilian tropical soils. A benchtop LIBS system was used for the analysis of 102 soil samples, prepared in the form of pressed pellets. Models for the prediction of clay, organic matter, pH, cation exchange capacity, base saturation, and the extractable nutrients P, K, Ca, and Mg were developed using univariate linear regression (ULR), multiple linear regression (MLR) and partial least squares regression (PLS). The following input data for PLS were used: (i) the full spectra from 200 to 540 nm (38,880 variables), and (ii) variables selected by the interval successive projections algorithm (iSPA). The multivariate models achieved satisfactory predictions [residual prediction deviation (RPD) > 1.40] for eight out of nine fertility attributes. However, the best performances were obtained for the PLS with the variable ranges selected by the iSPA, which achieved satisfactory predictions (RPD ≥ 1.44) for seven out of the nine soil attributes studied. The MLR method obtained lower prediction performance than the iSPA-PLS using only 21 variables. The iSPA-PLS approach allowed a reduction from 3 to 160-fold in the total of variables compared to the full LIBS spectra, making it efficient and accurate modelling method that uses reduced number of variables. Although LIBS technique proved to be efficient for predicting fertility attributes in tropical soils, further research is encouraged in order to reduce the amount of sample preparation conducted in this study.

Introduction

Worldwide, sensor-based analytical methods have been seen to characterize key soil fertility attributes in a rapid way without producing chemical residues (Viscarra Rossel and Bouma, 2016, Molin and Tavares, 2019). Proximal soil sensing (PSS) methods allows to reduce the number of operations con-ducted in traditional analytical procedures (e.g. digestion and extraction), making the analysis faster, cost-effective and environmentally friendly. Different PSS techniques have been applied for determining soil fertility attributes, e.g., X-ray fluorescence (XRF), mid infrared and visible and near infrared (vis-NIR) diffuse reflectance spectroscopies, ion selective electrodes (ISE), are few to mention among others (Silva and Molin, 2018, Munnaf et al., 2019, Demattê et al., 2019, Tavares et al., 2020a). LIBS spectra allow a broad characterization of soil elementary constitution (e.g., N, P, K, Ca, Mg, Al, and Fe) (He et al., 2018, Nicolodelli et al., 2019), and satisfactory determinations of the total contents of these elements in soil samples are commonly found in the literature (Hussain et al., 2007, Riebe et al., 2019, Erler et al., 2020, Xu et al., 2019a). LIBS spectra can also be used as a proxy to infer about soil fertility attributes [e.g., extractable (ex-) nutrients, pH, cation exchange capacity (CEC), organic matter (OM), pH, base saturation (V), and textural attributes] (Senesi, 2020). However, research regarding LIBS applications for soil fertility analysis are still incipient. In temperate soils, the results of the few studies carried out are not always satisfactory. For example, in German agricultural soils, Erler et al. (2020) have achieved good prediction performances for pH (R² between 0.91 and 0.95), reasonable predictions for humus content (R² between 0.54 and 0.66), and poor predictions for ex-P (R² between 0.22 and 0.35). Similarly, in Chinese agricultural soils, studies have shown poor prediction performance for ex-K and ex-P (Xu et al., 2019a), reasonable predictions for OM (R² = 0.58) (Xu et al., 2019b), and good prediction for CEC, OM, and pH (R² ≥ 0.81) (Xu et al., 2019a). Xu et al. (2019c) reported satisfactory predictions for pH, OM, and total N (R² between 0.60 and 0.75), and poor performances (R² < 0.35) for ex-K and ex-P in Chinese soils. In Brazilian tropical soils, although good prediction performance (R² ≥ 0.85) have been achieved for textural attributes (Villas-Boas et al., 2016) and pH (Ferreira et al., 2015), evaluation of the prediction accuracy for other key fertility attributes (e.g., CEC, V, OM, and extractable nutrients) has not been addressed in the literature yet.

The selection of an optimal modelling procedure for LIBS spectral data, capable of extracting useful but hidden information is crucial for accurate prediction of soil fertility attributes. Although there is no consensus regarding the best modelling strategy to adopt, recent studies suggested using multivariate techniques for the determination of soils elementary constitution when using LIBS spectra, in particular the partial least squares regression (PLS) (Takahashi and Thornton, 2017, Riebe et al., 2019). In addition, multivariate models are a simple and useful strategy to mitigate the matrix effect commonly present in soil samples (Takahashi and Thornton, 2017, Tavares et al., 2020b). Despite of the benefits, multivariate models using the entire spectrum require high computational capacity, especially due to the high spectral resolution of LIBS data that result in a large number of spectral variables (e.g., thousands data points) (Erler et al., 2020). In this regard, algorithms for the selection of most significant variables, such as deep learning or interval successive projection algorithm in partial least squares (iSPA-PLS) (Gomes et al., 2013, Riebe et al., 2019, Niu et al., 2021), can be a useful technique to reduce the amount of LIBS input variables and simplify data acquisition and processing procedures. To our best knowledge, no previous work reported the performance of iSPA-PLS for the prediction of key soil attributes in tropical soils, in comparison with full-spectra based PLS analyses.

This work aimed at evaluating different modelling approaches, namely, multiple linear regression (MLR) using selected emission lines of LIBS data, PLS using the full LIBS spectra from 200 to 540 nm (with 38,880 variables), and PLS using specific spectral regions selected by the iSPA algorithm for the prediction of key soil fertility attributes in Brazilian tropical soils. The best performing model was compared with univariate linear regression (ULR) in order to assess the influence of the matrix effect on the prediction of extractable nutrients (ex-P, ex-Ca, and ex-Mg).