为提供一种高效、 快速和无损的粳稻叶片氮素含量反演方法, 以粳稻小区试验为基础, 利用高光谱技术和室内化学实验, 获取粳稻分蘖期、 拔节期和抽穗期三个生育期共280组叶片高光谱数据以及相对应的水稻叶片氮素含量数据, 分析不同施氮水平的粳稻叶片光谱特征, 采用随机青蛙算法(random_frog)与迭代和保留信息变量算法(IRIV)相结合的方式筛选特征波段, 并将任意两个光谱波段随机组合构建差值植被指数DSI(Ri, Rj)、 比值植被指数RSI(Ri, Rj)和归一化植被指数NDSI(Ri, Rj), 分别将较优的特征波段组合和植被指数组合作为模型输入, 构建BP神经网络、 支持向量机（SVR）和非支配的精英策略遗传算法优化极限学习机（NSGA2-ELM）粳稻叶片氮素含量反演模型, 并对模型进行验证分析。 结果表明: 随着施氮水平的增加, 粳稻叶片近红外波段范围反射率逐渐升高, 在可见光波段范围反射率逐渐降低。 采用random_frog与IRIV相结合的方式筛选特征波段共得到8个特征波段, 其中可见光波段7个, 分别为414.2, 430.9, 439.6, 447.9, 682.7, 685.4和686.3 nm, 近红外波段仅有1个为999.1 nm, 该方法较好地剔除了干扰信息, 大大降低了波段间的共线性。 同时从三种植被指数（DSI(Ri, Rj), RSI(Ri, Rj), NDSI(Ri, Rj)）与粳稻叶片氮素含量的决定系数等势图中可知, DSI(R648.1, R738.1), RSI(R532.8, R677.3)和NDSI(R654.8, R532.9)与叶片氮素含量相关性最好, R2分别为0.811 4, 0.829 7和0.816 9。 在输入参量不同的建模效果对比分析中, 以特征波段组合作为模型输入所构建的模型反演效果略优于植被指数组合, R2均大于0.7, RMSE均小于0.57。 而在反演模型间的对比分析中, 提出的NSGA2-ELM反演模型的估测效果要优于BP神经网络模型和SVR模型, 训练集决定系数R2为0.817 2, 均方根误差RMSE为0.355 5, 验证集R2为0.849 7, RMSE为0.301 1。 鉴于此, random_frog-IRIV筛选特征波段方法结合NSGA2-ELM建模方法在快速检测粳稻叶片氮素含量中具有显著优势, 可为粳稻田间精准施肥提供了参考。

In order to provide an efficient, rapid and non-destructive Inversion method for the nitrogen content of japonica rice leaves, based on the japonica rice plot experiment, using high spectrum technology and laboratory chemistry experiments to obtain the effective data for three growth periods of japonica rice in the tillering stage, joining stage and heading stage. A total of 280 sets of leaf high spectrum data and corresponding rice leaf nitrogen content data were obtained to analyze the spectral characteristics of japonica rice leaves with different nitrogen treatment levels. The random frog algorithm (Random_frog) is combined with the iteratively retaining informative variables algorithm (IRIV) to screen the feature bands, and any two spectral bands are randomly combined to construct the difference vegetation index DSI (Ri, Rj), ratio vegetation index. RSI (Ri, Rj) and normalized vegetation index NDSI (Ri, Rj), respectively, combine the superior feature band combination and vegetation index as model inputs. Therefore, BP neural network, support vector machine (SVR) and non-dominated elite strategy genetic algorithm optimization limit learning machine (NSGA2-ELM) japonica rice leaf nitrogen content Inversion model were constructed, and the model was verified and analyzed. The results showed that with the increasing level of nitrogen fertilizer treatment, the reflectivity of the japonica rice leaves in the near-infrared range gradually increased, while the reflectance decreased gradually in the visible range. A total of 8 characteristic bands were obtained by combining Random_frog and IRIV. Among them, there are 7 visible light bands, which are 414.2, 430.9, 439.6, 447.9, 682.7, 685.4 and 686.3 nm, respectively. Only one of the near-infrared bands is 999.1 nm. This method effectively eliminates interference information and greatly reduces the collinearity between the bands. At the same time, it can be analyzed from the three vegetation index (DSI (Ri, Rj), RSI (Ri, Rj), NDSI (Ri, Rj)) and the determination coefficient of the nitrogen content of the japonica rice leaves. DSI (R648.1, R738.1), RSI (R532.8, R677.3) and NDSI (R654.8, R532.9) have the best correlation with leaf nitrogen content, R2 are 0.811 4, 0.829 7 and 0.816 9, respectively. In the comparative analysis of the input parameters with different input parameters, the model Inversion with the feature band combination as the model input is slightly better than the vegetation index combination, R2 is greater than 0.7, and the RMSE is less than 0.57. In the comparative analysis between the Inversion models, the estimation effect of the NSGA2-ELMInversion model proposed in this paper is significantly better than the BP neural network model and the SVR model. The training set determination coefficient R2 is 0.817 2, and the root means square error RMSE is 0.355 5. The set R2 is 0.849 7 and the RMSE is 0.301 1. According to the results of this study, the Random_frog-IRIV screening characteristic band method combined with NSGA2-ELM modeling method has a significant advantage in rapidly detecting the nitrogen content of japonica rice leaves. The research results can provide a theoretical reference for field precision fertilization of japonica rice.