基于CARS-RUN-ELM算法的水稻叶片氮磷含量协同反演方法

同时反演氮、磷元素含量相对于单一元素反演可以更加全面地表达水稻的营养状况，为快速、准确获取水稻叶片氮、磷含量和精准变量施肥提供依据。该研究基于不同氮肥处理的田间小区试验，获取水稻叶片氮、磷含量数据，采用竞争性自适应重加权采样法（Competitive Adapative Reweighted Sampling，CARS）筛选氮素与磷素共同特征波长，以特征波长反射率为输入，以化学方法测得叶片氮、磷元素含量为输出，分别使用反向传播神经网络、极限学习机（Extreme Learning Machine，ELM）、龙格-库塔算法优化极限学习机（RUNge Kutta optimizer-Extreme Learning Machine，RUN-ELM）构建水稻叶片氮、磷含量反演模型并分析。结果表明：采用CARS方法有效去除了高光谱中大量无用、冗余信息，得到5个氮、磷元素共同特征波长，去除具有共线性的特征波长，最后筛选出的特征波长分别是451、488、780、813 nm。使用筛选后的特征波长反射率构建RUN-ELM水稻叶片氮、磷含量反演模型效果最好，氮素训练集的决定系数R2为0.690，均方根误差为0.669 mg/g，磷素训练集的决定系数R2为0.620，均方根误差为0.027 mg/g。通过对比，RUN-ELM在预测能力、模型稳定性上优于反向传播神经网络以及ELM模型。综上研究，基于CARS-RUN-ELM的水稻叶片氮、磷含量反演模型可以快速、准确获取水稻叶片氮、磷含量，可为水稻精准施肥提供参考。

Simultaneous inversion method of nitrogen and phosphorus contents in rice leaves using CARS-RUN-ELM algorithm

Abstract: Chemical fertilizers can be the zero and negative growth for the high requirement from the environmental and green development in recent years. Precise fertilization on demand can depend mainly on the rapid and accurate detection of crop nutrition and health status in fields. Among them, a synergistic inversion of nitrogen and phosphorus content can be expected to more comprehensively express the nutritional conditions of rice, compared with the single element inversion. It is also of great significance to the rice field management and accurate fertilization at the greening and tillering stage. In this study, a series of field plot experiments were conducted to realize the different treatments of nitrogen fertilizer. A chemical experiment was selected to obtain the nitrogen and phosphorus content in rice leaves, while a marine optical fiber spectrometer was used for the hyperspectral data of rice leaves. The data sets of nitrogen content were then sorted after measurement. A Kolmogorov-Smirnov test was also utilized to randomly divide the data sets into the 224 training and 93 verification sets, according to the ratio of 7:3. Competitive Adaptive Reweighted Sampling (CARS) was then used to screen the common characteristic wavelengths of nitrogen and phosphorus from the data sets. As such, the reflectivity of characteristic wavelengths was set as the input, whereas, the measured contents of nitrogen and phosphorus in the rice leaves were used as the output. A Back Propagation (BP) neural network, Extreme Learning Machine (ELM), and Runge-Kutta optimizer-Extreme Learning Machine (RUN-ELM) were used to construct the inversion models of nitrogen and phosphorus contents in the rice leaves. The results show that the CARS effectively removed a large number of redundant information in the hyperspectra data, where five common characteristic wavelengths of nitrogen and phosphorus were obtained to remove the collinearity characteristic wavelengths. After that, the characteristic wavelengths were selected as 451, 488, 780, and 813 nm. The best performance of the RUN-ELM model was achieved to retrieve the nitrogen and phosphorus contents in the rice leaves using the selected reflectance of characteristic wavelength. The determination coefficient and Root Mean Square Error (RMSE) of the nitrogen training set were 0.690 and 0.669 mg/g, respectively, while the determination coefficient and RMSE of the phosphorus training set were 0.620 and 0.027 mg/g, respectively. By contrast, the RUN-ELM model was superior to the BP neural network and ELM model in the prediction and simulation. Furthermore, the higher accuracy and stability of the ELM model were realized to improve the better weight and threshold, where the local optimal solution was avoided for the higher convergence speed than before. The reason was that the promising region was searched in the space using the calculated slope as the search logic and the Enhanced Solution Quality (ESQ) mechanism, according to the calculating gradient during Runge-Kutta (RK) optimization. To sum up, the CARS-RUN-ELM inversion model can rapidly and accurately extract the nitrogen and phosphorus content in the rice leaves. The high accuracy and stability of the model can greatly contribute to effectively gaining the nutrient element contents of rice leaves. The finding can provide a strong reference to detect the nitrogen and phosphorus content for the precise fertilization of rice on demand.