Diffuse reflectance spectroscopy to estimate the concentration of chemical elements in soil and sediment combining pre-processing methods with machine learning

The intense soil use with inadequate management can result in the constant transport of sediments with chemical elements absorbed to aquatic systems. The diffuse reflectance spectroscopy in the near infrared (NIR) and medium (MIR) spectral bands associated with chemometry and machine learning, is an analytical technique that has the potential to quantify the concentration of chemical elements in the environment. However, there is no consensus on the best combination of calibration methods, spectral pre-processing and spectral ranges. Thus, the objective of this study was to evaluate the use of this technique, with the combination of different spectral bands, pre-processing techniques and machine learning to estimate the concentration of chemical elements on soil and sediment samples. In this study we used a soil and sediment database from samples collected in the Guaporé River catchment, in southern Brazil. A total of 316 soil samples and 196 sediment samples were dried, disaggregated and sieved at 63 μm. Organic carbon (CO) was quantified by wet oxidation and the total concentration of 21 elements (Al, Ba, Be, Ca, Co, Cr, Cu, Fe, K, La, Li, Mg, Mn, Na, Ni, P, Pb, Sr, Ti V and Zn) were quantified by ICP-OES after microwave assisted digestion for 9,5 min at 182ºC with HCl and HNO₃ concentrated in the proportion of 3:1. The NIR (1000-2500 nm) and MIR (2500-25000 nm) spectra were obtained in all soil and sediment samples. Two machine-learning methods were tested: Partial Least Squares Regression (PLSR) and Support Vector Machine (SVM), associated with three different spectrum pre-processing methods: Detrend (DET), Savitzky-Golay Derivative (SGD) and Standard Normal Variate (SNV), compared to raw data (RAW). Performance was assessed by the coefficient of determination (R²) and the relationship between performance and interquartile distance (RPIQ). The SVM model resulted in better predictions compared to the PLSR in all evaluated cases, as indicated by the average adjustment values of the model (R²=0.87 for SVM and 0.62 for PLSR), and by the RPIQ values (7.14 for SVM and 2.22 for PLSR). The pre-processing method increased the accuracy of the estimates in the following order: RAW<SNV< DET<SGD. The best performance in relation to the spectral range was observed for the MIR region, being significantly superior to the NIR and NIR+MIR combination. The adjustment of the models calibrated with soil (R²=0.91) and sediment (R²=0.90) data was higher compared to the calibrated with the combination soil + sediment (R²=0.78). For RPIQ, the calibration model with soil data showed the highest RPIQ value (9.29), being higher and differing significantly from the others. In general, the results show that the combination of different calibration methods, spectral pre-processing and spectral ranges has an effect on the accuracy of the estimates. The studied elements can be estimated by means of diffuse reflectance spectroscopy, however it should be noted that this technique has an associated error in the estimates due to the heterogeneity of the chemical structure of the elements in the soil and sediment matrix and the reference samples obtained by chemical methods.