Prediction of soil attributes using the Chinese soil spectral library and standardized spectra recorded at field conditions

Abstract

Organic matter (OM), total nitrogen (TN), and pH are essential soil properties for assessing the fertility of paddy soils. They can be measured with visible and near infrared (vis-NIR) spectroscopy effectively in the field. However, environmental factors e.g., soil moisture and particle size distribution affect the accuracy of spectroscopic measurement and successful calibration transfer between laboratory and field spectra. Large spectral libraries derived from dried and ground soil samples thus could not be used to predict soil properties using spectra of fresh (non-processed) samples. In this paper, we investigated the possibility of using the Chinese soil spectral library (CSSL) of dry ground soils to predict OM, TN and pH of paddy soils in the Yangtze River Delta using spectra of fresh (non-processed) soil samples measured in situ, after removing the influences of the environmental factors with direct standardization (DS). The locally weighted regression (LWR) model built on the CSSL was then used to predict with the DS-transferred field spectra. The CSSL consists of vis-NIR spectra of over 3993 samples collected local dataset from 19 Chinese provinces. Two hundred and twenty-five soil samples independent from the CSSL (local dataset) were collected from 20 target sites in the Yangtze River Delta, China and their spectra were measured in both field and laboratory conditions. Using DS, a subset of the corresponding field and laboratory spectra from the independent set (designated as the transfer set) was used to derive the DS transfer matrix, which characterized the differences between the field and laboratory spectra. The field spectra of the 225 samples were then transferred to match characteristics of laboratory measured spectra of processed soil samples. The predictions of soil properties were performed on the DS-transferred field spectra using a LWR model derived with the CSSL. Results showed that DS effectively removed the effects of moisture from field spectra, and led to simultaneous improvement in the predictions of pH, OM, and TN to an acceptable level (pH: R2 = 0.611, root mean square error (RMSE) = 0.73 and ratio of performance to inter-quartile range (RPIQ)= 2.30; OM: R2 = 0.641, RMSE=6.82gkg-1 and RPIQ = 1.79; TN: R2 = 0.658, RMSE = 0.39gkg-1 and RPIQ = 1.81). We recommended the use of DS combined with CSSL models for the efficient prediction of soil pH, OM, and TN simultaneously using field scans of paddy soils.