Laboratory assessment of three quantitative methods for estimating the organic matter content of soils in China based on visible/near-infrared reflectance spectra

Abstract Hyperspectral reflectance data (350 nm to 2500 nm) were recorded for five different soil types originating from seven eco-climatic zones in middle and eastern China. Three spectral formats were prepared for use in the development of soil organic matter content (SOM) prediction models: (1) original spectral reflectance (OR), (2) first derivative spectra corrected using the Savitzky–Golay (FD-SG) technique and (3) using Norris (FD-NG) smoothing filters. All two-band combinations of the three types of spectra in ratio index (RI), normalized difference index (NDI) and difference index (DI) were used in linear and non-linear regression analyses with SOM content. A new difference index [DI(NDR554, NDR1398)] composed of the first derivative spectra at 554 nm and 1398 nm with a FD-NG smoothing filter gave the best prediction of SOM in all two-band indices. Partial least square (PLS) models using FD-NG and calibration of the spectral regions 500–900 nm and 1350–1460 nm resulted in an R2 of 0.91 (n = 331), indicating better performance than that obtained using OR and FD-SG and the entire spectral region (400–2500 nm) or their sensitive bands. With principal components extracted by PLS serving as inputs for a backpropagation neural network (BPNN), the PLS–BPNN calibration model for estimating SOM resulted in an improved performance (R2 = 0.98; n = 331). The results were tested with independent validation datasets, which indicated that the models were reliable for SOM estimation with predictive accuracy in the following sequence: PLS–BPNN model > PLS > DI(NDR554, NDR1398). Thus, the PLS–BPNN model may serve as a useful tool for estimating SOM content with high prediction accuracy. Since DI(NDR554, NDR1398) needed only two derivative spectral bands, it is also recommended as a good spectral index for reliably estimating SOM. These results are of significant potential valuable in the field of soil ecosystem observation, specifically for sensor selection and future portable sensor design. Although they were shown to be useful in the five different soil types from middle and eastern China, these models and methods should be further tested in soils sampled from other regions and countries to prove their validity and applications.