Selection of "Local" Models for Prediction of Soil Organic Matter Using a Regional Soil Vis-NIR Spectral Library. Issue 1 (January 2016)

• Record Type: Journal Article
• Authors: Zeng, Rong; Zhao, Yu-Guo; Li, De-Cheng; Wu, Deng-Wei; Wei, Chang-Long; Zhang, Gan-Lin
• Is Part Of: Soil science. Volume 181:Issue 1(2016)
• Publication Date: 2016-01
• Subjects: SOM contents -- spectroscopy -- spectral library -- Vis-NIR; Soil science -- Periodicals; 631.405
• DOI: 10.1097/SS.0000000000000132
• Languages: English
• ISSNs: 0038-075X

Abstract

Abstract : Abstract: Soil spectral libraries have been established as a reference for predicting soil properties by visible and near-infrared (Vis-NIR) spectroscopy. Numerous studies show that predictions of soil properties over a local area can be improved by selecting an appropriate "local" subset from a large library; although these have usually been geographically local, they can be local in other than the geographic sense. We investigated prediction of soil organic matter at a local site using a regional soil Vis-NIR spectral library with 1, 365 samples. Models built using the entire library were compared with subsets selected by (i) parent material (Cali_pm), (ii) land use type (Cali_lu), (iii) material-land use combination (Cali_com), and (iv) spectral similarity (Cali_ss). Models were built by partial least-squares regression, and their performances were evaluated using two independent test sets, one for paddy field (Test_paddy) and another one for upland agriculture (Test_up). Prediction accuracy was measured by the ratio of percentage deviation (RPD) compared with models built on the entire library. Ratios of percentage deviation for Cali_lu increased from 1.58 to 1.65 (Test_up) and from 2.05 to 3.02 (Test_paddy); for Cali_ss, RPD increased from 1.58 to 1.89 (Test_up) and from 2.05 to 2.26 (Test_paddy). Cali_pm models performed well for Test_paddy (RPD = 2.76) but poorly for Test_up (RPD = 1.11). Cali_com models used the fewest number of samples and performed poorly for both test sets (RPD < 1.5). These results show the potential of using land use types or spectral similarity to select "local" models for prediction of soil organic matter using a regional spectral library. Abstract : Supplemental digital content is available in the text.