Coupling discrete wavelet packet transformation and local correlation maximization improving prediction accuracy of soil organic carbon based on hyperspectral reflectance

Soil organic carbon (SOC) is an essential soil property for assessing the fertility of paddy soils. It can be measured with visible and near infrared spectroscopy effectively in the field. Meanwhile, there are a lot of factors, such as soil water, surface conditions and so on, which might affect the spectra, increasing the difficulty in extracting the effective information, and reducing the prediction accuracy of SOC content. Noise reduction must be considered in developing hyperspectral estimation models, but how to reduce noise while retaining as much useful information as possible needs for investigation. As advanced spectral mining methods, local correlation maximization (LCM) arithmetic was used to solve this problem in this study. In the present study, a total of 70 soil samples of paddy soil were collected from rice fields in Zhulin town, Jintian city, Jiangsu Province. The sample holders were clear aluminum boxes in 7 cm diameter and 3 cm deep, which were filled and leveled at the rim with a spatula. Reflectance of soil samples measured using ASD Fieldspec 3 Spectrometer in a dark room when brought these samples indoor immediately to keep them in the field conditions. We used the following steps to process soil reflectance: First, discrete wavelet packet transformation (DWPT) was used to decompose the original spectral (result from 0.6-order differential) in 7 levels using Bior1.3 wavelet basis by MATLAB programming language. In order to select the maximum level of DWPT, correlation coefficients between SOC and the spectra of each level was computed. Secondly, LCM method was used to develop the local optimal correlation coefficient (LOCC) and optimal band which was determined from the optimal correlative curve and the optimal spectra (OS), respectively. Thirdly, a PLSR model was built to predict SOC contents. And then, determination coefficient of validation (R2 v), root mean square error of validation (RMSEV), and residual prediction deviation (RPD) were used for accuracy assessment. We also used variable in the projection (VIP) analysis to identify the reason why LCM could improve the accuracy of predict model at the same time. The results showed: 1) significant correlated bands followed increasing-decreasing trend with the increase of wavelet decomposed level and the maximum level identified as level 6. This implied that the wavelet packet transformation amplified some useful SOC information that was previously obscured by noise. 2) optimal spectra that established from LCM could effectively remove noise while preserving the detail information of SOC simultaneously. 3) compared with raw spectral (R2 v=0.693, RMSEV=1.952 g/kg, RPD=1.85), the wavelet packet transformation provided good results (R2 v=0.727, RMSEV=1.840 g/kg, RPD=1.97) of SOC prediction, combined with LCM arithmetic, the model had the best performance (R2 v=0.781, RMSEV=1.679 g/kg, RPD=2.17) to predict SOC content. According to VIP score, important bands for SOC prediction hadthree pink values, two of them located in the characteristic bands of soil water, this illustrated LCM can't remove the effects of soil water thorough. Results indicated that the discrete wavelet packet transformation and local correlation maximization (DWPT-LCM) method had great potential to monitor SOC contents in paddy soils when reduced white noise while retaining as much soil organic carbon information as possible. © 2017, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.