Estimating Canopy Nitrogen Content of Rice Using Hyperspectral Reflectance Combined with SG-FD-CARS-ELM in Cold Region

In this study, visible and near infrared hyperspectral imaging technique was used to predict canopy leaf nitrogen content (CLNC) of rice in cold region. Canopy hyperspectral images of rice were acquired at tillering, jointing and heading stage, respectively. Original spectra was extracted using ENVI5.0 software, and leaf nitrogen content was obtained by chemical analysis method. 5 pre-processing methods of savitzky-golay smoothing (SG), multiplicative scatter correction (MSC), standard normal variate (SNV), first derivative (FD) and second derivative (SD) were used to eliminate unexpected noise. After comparing the performance of PLSR models based on spectra of full wavelengths after pre-processing, SG combined with FD had the best performance for eliminating the noise interference and improving the performance of models. In order to further simplify and enhance the models, 3 variable selection methods of successive projections algorithm (SPA), uninformative variable elimination (UVE) and competitive adaptive reweighted sampling (CARS) were used to select the characteristic wavelengths, and partial least square regression (PLSR) and extreme learning machine (ELM) were used to establish prediction models. After comparing the performance of PLSR models and ELM models, CARS could effectively select the wavelengths that had strong information and were not sensitive to external disturbance factors, and the nonlinear ELM model was more suitable for predicting CLNC of rice in cold region, the specific values of RC 2 and RP 2 of ELM models based on CARS were 0.906 and 0.888 for tillering stage, 0.903 and 0.892 for jointing stage, and 0.894, 0.887 for heading stage, respectively. The results of this study could provide a reference for quantitative analysis of nitrogen content of rice using hyperspectral technology.