Optimization method of calibration dataset for VIS-NIR spectral inversion model of soil organic matter content

Abstract: Soil organic matter (SOM) is not only an important indicator of soil fertility but also an important source and sink of the global carbon cycle. Therefore, it is essential to acquire the information of SOM for soil management. Visible and near-infrared (VIS-NIR) reflectance spectroscopy, known as a novel, rapid, accurate, environment-friendly and efficient approach compared with conventional laboratory analyses, has proven to be promising in the acquisition of various soil properties. Construction of a calibration set is key to the VIS-NIR quantitative analysis in building up a prediction model of high quality. The aim of this paper was to explore how the sample selection method and the number of samples may affect the accuracy of VIS-NIR models for SOM estimation. A total of 100 paddy soil samples (0-15 cm) were collected from the Honghu City, which is located in the Jianghan Plain, China. After air drying, grinding and sieving (0.25 mm), reflectance of these pretreated samples was measured with FieldSpec3 (Analytical Spectral Devices Inc., America). Three samples were neglected after outlier detections of spectra and SOM content. Out of the remaining 97 samples, 20 samples were selected by means of concentration gradient, which then formed the validation sample set. The remaining 77 samples formed the total calibration set. With SOM content or soil spectral information as inputs, 3 sample selection methods, namely Kennard-Stone (KS), sample set partitioning based on joint X-Y distance (SPXY) and Rank-KS, were used in the construction of calibration subsets with different proportions of the samples in total calibration set, such as 10% and 20%. Based on the different calibration subsets, partial least squares regression (PLSR) was used for model calibrations. Results showed that the calibration set selected by KS approach could not improve model predictive capability compared with the total calibration set. The KS approach, however, could reduce as many as 30% samples of the total calibration set while the ratio of performance to standard deviation (RPD) was retained above 2.0. The SPXY approach performed the best when 50% samples of the total calibration set were selected in the model calibration. The determination coefficient for calibration (Rc2) reached 0.922, the determination coefficient for prediction (Rp2) was 0.848, and the RPD reached 2.557. This was because the SPXY approach took into account both SOM content and soil spectra in the sample selection process. With only 30% samples of the total calibration set selected by the Rank-KS method, it had the lowest cost of calibration with satisfactory performance (Rc2=0.872, Rp2=0.802 and RPD=2.212). Overall, such results indicate that it is possible to reduce the number of calibration samples while retaining or even improving the predictive capacity of VIS-NIR models for SOM estimation. All the 3 calibration selection approaches have been proven to be useful for the improvement of model practicability.