关键词: 养分, 高光谱, 预处理, 降维

Abstract:

Plant nutrient status is a comprehensive response to soil nutrient supply, crop nutrient demand, and crop nutrient abilities. Detecting variations in plant nutrient content is an important aspect of forest management. However, conventional chemical analysis techniques are often time and labor intensive, particularly when applied over large areas. In recent years, some convenient and non-destructive tools have been applied to monitor plant biochemical properties; however, there is no agreement about which methods are most reliable. Among the available methods, some employ hyperspectral data to nondestructively estimate levels of nitrogen, phosphorus, and potassium in plants, thus providing a theoretical framework to support scientific forest management. Certain optical characteristics in the visible and near-infrared regions are closely associated with the absorption features of chlorophyll, other pigments, water, and chemicals in leaves and canopies. However, the efficacy of utilizing spectral data to detect various nutrient parameters is dependent on the data processing methods employed. In this study, we applied near-infrared spectroscopy to examine the leaves of nineteen wetland forest species and assessed various models’ performances in estimating Total Nitrogen (TN), Total Phosphorus (TP) and Total Potassium (TK) content in the vegetation. Eleven spectral preprocessing methods and three spectral data dimensionality reduction methods were used to preprocess the spectra. And two of algorithms, the Partial Least Squares Regression (PLSR) and Support Vector Machine Regression (SVR), were used to develop the nutrients prediction models. The determination coefficients (R ²) and Root Mean Square Error (RMSE) of the models were used to evaluate the performance of the models for calibration, cross validation and prediction datasets. The Relative Percent Difference (RPD) for the prediction dataset was also used to assess the models. Results showed that the Standard Normal Variate (SNV) approach combined with the first derivative (1 ^st) preprocessing method had the highest accuracy among the 11 data pretreatment approaches, with RPD values of 2.35, 2.39, and 2.45 for TN, TP, and TK, respectively. Among the different dimensional-reduction methods, the Principal Component Analysis (PCA) performed the best, and SVR outperformed PLSR in parameter estimation. Models incorporating the SVR algorithm and data preprocessed using the SNV+1 ^st approach yielded the best prediction results for the three parameters. The best model for TN had ${R^{2}}_{p}$, RMSEp, and RPD values of 0.85, 2.82% and 2.50, respectively; best model for TP had ${R^{2}}_{p}$, RMSEp, and RPD values of 0.90, 0.55%, and 2.83, respectively; and best model for TK had ${R^{2}}_{p}$, RMSEp, and RFD values of 0.85, 3.80%, and 2.60, respectively. The results indicated that visible and near-infrared spectra can be used to estimate the leaf TN, TP, and TK content of wetland trees. However, before model calibration, the proper preprocessing of the spectral data is necessary to improve the performance of the models.

Key words: nutrient, hyperspectral, pretreatment, dimension reduction