Modelling soil water retention and water‐holding capacity with visible–near‐infrared spectra and machine learning. (1st March 2022)

• Record Type: Journal Article
• Authors: Baumann, Philipp; Lee, Juhwan; Behrens, Thorsten; Biswas, Asim; Six, Johan; McLachlan, Gordon; Viscarra Rossel, Raphael A.
• Is Part Of: European journal of soil science. Volume 73:Number 2(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-03-01
• Subjects: available soil water -- machine learning -- soil water retention -- visible–near‐infrared spectroscopy -- water retention models; Soil science -- Periodicals; 631.4
• DOI: 10.1111/ejss.13220
• Languages: English
• ISSNs: 1351-0754

Abstract

Abstract: We need measurements of soil water retention (SWR) and available water capacity (AWC) to assess and model soil functions, but methods are time‐consuming and expensive. Our aim here was to investigate the modelling of AWC and SWR with visible–near‐infrared spectra (vis–NIR) and the machine‐learning method cubist . We used soils from 54 locations across Australian agricultural regions, from three depths: 0–15 cm, 15–30 cm and 30–60 cm. The volumetric water content of the samples and their vis–NIR spectra were measured at seven matric potentials from −1 kPa to −1500 kPa. We modelled the following: (i) AWC directly with the average spectra of the samples measured at different water contents, (ii) water contents at field capacity and permanent wilting point and calculated AWC from those estimates, (iii) AWC with spectra of air‐dried soils, and (iv) parameters of the Kosugi and van Genuchten SWR models, then reconstructed the SWR curves to calculate AWC. We compared the estimates with those from a local pedotransfer function (PTF) and an established Australian PTF. The accuracy of the spectroscopic approaches varied but was generally better than the PTFs. The spectroscopic methods are also more practical because they do not require additional soil properties for the modelling. The root‐mean squared‐error (RMSE) of the spectroscopic methods ranged from 0.033 cm 3 cm −3 to 0.059 cm 3 cm −3 . The RMSEs of the PTFs were 0.050 cm 3 cm −3 for the local and 0.077 cm 3 cm −3 for the general PTF. Spectroscopy with machine learning provides a rapid and versatile method for estimating the AWC and SWR characteristics of diverse agricultural soils. Highlights: Soil available water capacity can be estimated with vis‐NIR specta. Parameters of water retention models can be estimated with vis‐NIR spectra. vis‐NIR spectroscopy performed better than pedotransfer functions. The results apply to a diverse range of soils.