为了通过植被指数（VI）准确、 可靠的获取不同施肥梯度、 不同品种的水稻叶面积指数（LAI）, 提出了一种基于改进的QGA-ELM算法应用于水稻LAI反演。 首先通过8折交叉验证确定极限学习机(ELM)最佳的隐含层神经元个数与隐含层激活函数类型, 再通过引入组合动态旋转角策略、 单点混沌交叉操作、 混沌变异操作、 确定性选择策略、 量子灾变操作对量子遗传算法（QGA）进行改进, 最后使用改进后的QGA算法优化ELM神经网络输入层到隐含层的连接权值和隐含层的阈值。 为了验证该模型普适性和有效性, 依次建立多元线性回归、 BP、 ELM、 QGA-ELM、 改进的QGA-ELM算法5种模型, 并在不同数据集上进行反演效果比较, 结果表明: （1）对比QGA-ELM算法和改进的QGA-ELM算法进化过程, 改进的算法能有效提升模型寻优能力, 避免算法早熟, 且能寻得更优结果。 （2）对比五种算法在不同数据集上的反演效果, 验证了NDVI, RVI与LAI之间主要为非线性关系, 且ELM神经网络模型反演效果要优于BP神经网络模型和多元线性回归模型。 （3）对比五种算法在不同数据集上的反演效果, 改进的QGA-ELM算法绝大部分情况下拥有最高的反演精度和最低的误差, 改进后的算法反演精度得到了明显提升, 泛化性能也得到了增强。 （4）改进的QGA-ELM算法在各种施肥梯度上均具有最高反演精度和最低误差, 且精度较高, 能为不同生长状况水稻LAI反演提供依据。 （5）五种模型对庆和香LAI反演精度均要高于龙稻18, 而改进的QGA-ELM算法在不同水稻品种上依然具有较高的反演精度, 且在不同水稻品种上反演精度相差极小, 远低于其他四种模型, 能很好适应不同水稻品种LAI反演要求, 极大提升模型的稳定性性, 为不同水稻品种反演提供参考意义。

In order to accurately and reliably obtain LAI of rice of different fertilization gradients and varieties through vegetation index (VI), an improved QGA-ELM algorithm was proposed in this paper for LAI inversion of rice. This model firstly determined by 8 fold cross-validation extreme learning machine (ELM) optimal number of neurons in the hidden layer and hidden layer activation function types, and by introducing a dynamic rotation Angle combination strategy, single point chaos crossover operation and mutation operation, deterministic selection strategy, quantum catastrophe operations to improve the quantum genetic algorithm (QGA), finally using the improved QGA ELM algorithm optimization neural network input layer to hidden layer connection weights and threshold of the hidden layer. In order to validate the model, this paper, in turn, to establish multiple linear regression, BP, ELM, QGA-ELM, improved QGA-five ELM algorithm model, and compared the inversion effect on different data sets, the results show that: (1) Compare the QGA-ELM evolution algorithm and the improved QGA-ELM algorithm, in this paper, the improved algorithm can enhance the searching capability model and avoid precocious, algorithm and can find better results. (2) By comparing the inversion effects of five algorithms on different data sets, it is verified that the relationship between NDVI, RVI and LAI is mainly non-linear, and the inversion effect of ELM neural network model is better than that of BP neural network model and multiple linear regression model. (3) By comparing the inversion effects of the five algorithms on different data sets, the improved QGA-ELM algorithm in this paper has the highest inversion accuracy and the lowest error in most cases, and the improved algorithm has significantly improved the inversion accuracy and generalization performance. (4) The improved QGA-ELM algorithm has the highest inversion accuracy and the lowest error in all fertilization gradients, and the accuracy is higher, which can provide a basis for LAI inversion of rice under different growth conditions. (5) Five model for Qinghexiang LAI inversion precision are higher than the dragon rice 18, and the improved QGA-ELM algorithm on different rice varieties still has high inversion accuracy, and the inversion precision tiny difference on different rice varieties, is far lower than the other four kinds of models, can adapt to different rice varieties LAI inversion requirements, greatly improve model stability, provide a reference for different rice varieties of inversion.