Determination of in-situ salinized soil moisture content from visible-near infrared (VIS–NIR) spectroscopy by fractional order derivative and spectral variable selection algorithms

Abstract

Abstract: The measurement of salinized soil moisture content (SSMC) is essential to precise irrigation and avoidance of secondary salinization. Visible and near infrared (VIS–NIR) spectroscopy has been effectively used to estimate soil moisture content (SMC) but not for SSMC. The direct application of in-situ VIS–NIR spectroscopy to the estimation of SSMC can help save a large amount of time and labor, but the in-situ VIS–NIR was interfered by many factors, such as soil texture, soil surface debris and environmental temperature. Spectral derivatives can be used to eliminate unnecessary interference for optimal spectral information, but traditional integer derivatives (i.e. first and second derivatives) often ignored some spectral information due to different integer order differential curves were obviously different. In addition, the full spectrum usually contains redundant spectral variables. These variables would affect the accuracy and estimation velocity of the model. Different combinations of fractional order derivative (FOD) and spectral variable selection algorithms (i.e. variable importance projection (VIP), competitive adaptive weighted sampling (CARS) and random frog algorithm (RFA)) may offer some alternative solutions to these problems. In order to test the effects of these combinations on VIS–NIR spectral model optimization, we measured the in-situ soil spectra of 163 sites in Shahaoqu Irrigation Area, Inner Mongolia, China. Meanwhile, we collected soil samples and measured their SSMC and soil salt content (SSC). Then the Extreme Learning Machine (ELM) model was applied to the SSMC estimation. The results showed that SSC and SSMC had obvious effects on in-situ spectra. With the increase of differential order, the spectral resolution increased gradually, but the spectral intensity decreased at the same time. So, the spectral information may not increase. However, FOD can balance the contradiction between spectral resolution and spectral intensity. The estimation of ELM models based on 0.75 order derivatives that is the most accurate among the full spectrum ELM models. The coefficient of determination ( R 2 ) was 0.83 and ratio of the performance to deviation (RPD) was 2.44. In all the models (twenty-seven different combinations of FOD and variable selection algorithms), the best model was based on the combination of 0.75 derivative spectrum and random frog algorithm ( R 2 = 0.94, RPD = 3.80). The results of this study also confirmed that the combination of RFA and FOD could effectively improve the accuracy of the in-situ spectral estimation of SSMC. However, VIP was chosen as an alternative due to computational efficiency. Keywords: salinized soil moisture content, in-situ visible and near-infrared spectroscopy, fractional order derivative, random frog algorithm, extreme learning machine DOI: 10.33440/j.ijpaa.20200303.98  Citation: Lao C C, Zhang Z T, Chen J Y, Chen H R, Yao Z H, Xing Z, Tai X, Ning J F, Chen Y W. Determination of in-situ salinized soil moisture content from visible-near infrared (VIS–NIR) spectroscopy by fractional order derivative and spectral variable selection algorithms.  Int J Precis Agric Aviat, 2020; 3(3): 21 –34.