关键词: 高光谱, 辣椒, 氮素含量, 机器学习, 连续投影算法, 径向基, 光谱分析

Abstract:

【Objective】Nitrogen is one of the essential nutrients for plant growth and development, and it plays an important role in strengthening chlorophyll synthesis in crops, enhancing plant resistance, and improving yield and quality. This study harnessed hyperspectral technology to swiftly, precisely, and non-invasively monitor nitrogen levels in pepper foliage throughout its growth cycle, delving into the correlation between leaf nitrogen content (LNC) and spectral reflectance characteristics. 【Method】The study was based on the hyperspectral data of pepper leaves collected from Guanzhuang Demonstration Base in Pepper Research Institute of Guizhou Academy of Agricultural Sciences in 2021. The research encompassed four pepper varieties (Qianjiao No. 8, Hongla No. 18, Layan 101, and Hong Global) and five different nitrogen fertilizer application rates (0, 120, 240, 360, and 480 kg·hm^-2). The pepper leaf spectral data were processed, involving Multiple Scatter Correction (MSC), Savitzky-Golay (SG) and First Derivative (FD), followed by the selection of sensitive bands using Pearson correlation coefficient, Successive Projections Algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS). Subsequently, three machine learning algorithms, such as Partial Least Squares Regression (PLSR), Random Forest (RF) and Radial Basis Function Neural Network (RBFNN), were employed to construct models for monitoring nitrogen levels in pepper leaves, to achieve the goals of enhancing agricultural production efficiency and accuracy, and realizing intelligent management and precise fertilization. 【Result】After preprocessing, the original spectra improved correlation coefficients significantly. Among these, the spectral data's inversion performance was notably superior after SG processing, with the effectiveness ranking as SG>FD>MSC>original spectra. Contrasting various band selection methods, the employing Pearson correlation coefficient for band selection resulted in bands being overly concentrated, leading to either redundant information or incomplete information extraction. While CARS algorithm selected bands across a broad range and in large quantities, its effectiveness was inferior to SPA due to containing more redundant information and noise. SPA-selected nitrogen content characteristic bands effectively reduced collinearity and redundant information, yielding the optimal model with the highest R² and the smallest RMSE. The performance of different modeling methods for pepper LNC estimation was as follows: RBFNN performed the best, followed by PLSR, with RF exhibiting the poorest performance. Among these, the SG-SPA-RBFNN combined model demonstrated the best inversion accuracy, with modeling results of R² =0.98 and RMSE =0.62, and validation results of R² =0.98 and RMSE =1.21, with an RPD of 3.08. RBFNN model excelled in handling high-dimensional spectral data, surpassing traditional PLSR and RF models. 【Conclusion】The hyperspectral reflectance characteristics were utilized to establish nitrogen content prediction models, which could effectively monitor nitrogen levels in pepper leaves, thereby enhancing agricultural management efficiency and providing the technical support for precise management and variable fertilization in pepper cultivation.

Key words: hyperspectral, chili peppers, nitrogen, machine learning, SPA, RBFNN, spectrum analysis