A Vis‐NIR Spectral Library to Predict Clay in Australian Cotton Growing Soil. Issue 6 (21st November 2018)

• Record Type: Journal Article
• Authors: Zhao, Dongxue; Zhao, Xueyu; Khongnawang, Tibet; Arshad, Maryem; Triantafilis, John
• Is Part Of: Soil Science Society of America Journal. Volume 82:Issue 6(2018)
• Page Start: 1347
• Page End: 1357
• Publication Date: 2018-11-21
• Subjects: Soils -- United States -- Periodicals; Soil science -- Periodicals; Periodicals; 631.4973
• DOI: 10.2136/sssaj2018.03.0100
• Languages: English
• ISSNs: 0361-5995

Abstract

Abstract : A vis‐NIR spectral libraries were built across seven cotton growing areas of Australia. Establishing a calibration from each area independently was more accurate than making a calibration from the other six areas. Model performance improved using a spiking algorithm. Model performance improved by combining topsoil and subsurface samples. To maintain profitability of cotton growing areas of Australia, information of nutrient management and water‐use efficiency are needed. In this regard, information about clay is required. This is a time‐consuming and expensive laboratory analysis to undertake. An alternative is to use visible‐near infrared (vis‐NIR) spectroscopy, which has shown potential at different scales (e.g., local and global). Here, we predicted clay using a machine learning algorithm (Cubist) from vis‐NIR acquired from topsoil (0–0.3 m) and subsurface (0.3–0.6 m) in seven cotton growing areas. The first aim was to assess the ability of soil samples from each area to predict clay independently. The second aim was to determine the ability of the samples of six areas to predict clay in an area withheld from the calibration. The third aim was to explore the potential to improve prediction using "spiking". The fourth was to determine how much data was necessary to establish a suitable library. We conclude that establishing a calibration from each area independently was more accurate than making a calibration from six areas and predicting clay from the area withheld from the calibration. We also found that improvements in model performance were possible using spiking. When using samples from topsoil or subsurface only, over 93 samples were required to obtain an accurate library. We also conclude that a combined dataset from topsoil and subsurface samples enabled a more consistent set of data with no loss of calibration and prediction accuracy, especially when considering the availability of calibration samples.