基于DWT-DE变换和AHA-ELM算法的水稻叶片氮含量预测方法

Prediction of Nitrogen Content in Rice Leaves Based on DWT-DE Transformation and AHA-ELM Algorithm

摘要:

为提供一种利用光谱数据对水稻氮素含量加以快速、无损、准确预测的方法，本文以东北水稻为研究对象，采集水稻3个生育期的高光谱数据，结合室内化学实验，为了提高氮素预测精度和模型可解释性，建立水稻氮素含量反演模型。将获取的高光谱数据和相对应的水稻叶片氮素含量，首先通过低通滤波方法对光谱数据进行预处理，针对处理后光谱数据，采用耦合离散小波和一阶微分变换（DWT-DE变换）对光谱数据进行降维，并分别与主成分分析（PCA）、离散小波多尺度分解方法进行对比。以降维后的结果作为输入，实测叶片氮素含量为输出，分别建立极限学习机（ELM）、粒子群优化支持向量机（PSO-SVM）和人工蜂鸟算法优化的极限学习机（AHA-ELM）反演模型，对水稻叶片氮素含量进行预测和验证。结果表明，采用耦合离散小波和一阶微分变换结果建立的AHA-ELM模型预测精度最高，预测效果优于ELM和PSO-SVM模型，训练集决定系数R2为0.8064，RMSE为0.3251mg/g，验证集R2为0.7915，RMSE为0.3620mg/g。鉴于此，本文提出的经DWT-DE变换建立的AHA-ELM模型在快速检测水稻氮素含量中有显著优势，可为水稻精准变量施肥提供参考。

关键词:水稻;氮素含量;高光谱;耦合离散小波和一阶微分变换;人工蜂鸟优化;极限学习机

Abstract:

In order to provide a rapid, non-destructive, and accurate prediction method for nitrogen content in rice using spectral data, focusing on northeast rice as the research object, hyperspectral data of rice in three growth stages were collected, and combined with indoor chemical experiments, aiming to improve the prediction accuracy and model interpretability of nitrogen content by establishing an inversion model for rice nitrogen content. The acquired hyperspectral data and corresponding nitrogen content of rice leaves were firstly preprocessed by using a low-pass filtering method. For the processed spectral data, a coupling discrete wavelet transform and first-order differential transform (DWT-DE transform) were used for dimensionality reduction, and compared with principal component analysis (PCA) and discrete wavelet multiresolution decomposition methods. The dimensionality-reduced results were used as inputs, and the measured leaf nitrogen content was the output, to establish inversion models by using extreme learning machine (ELM), particle swarm optimization support vector machine (PSO-SVM), and artificial hummingbird algorithm optimized extreme learning machine (AHA-ELM), respectively, for predicting and validating rice leaf nitrogen content. The results showed that the AHA-ELM model established using the results of the coupling discrete wavelet and first-order differential transform had the highest prediction accuracy, which was superior to the ELM and PSO-SVM models. The determination coefficient R2 of the training set was 0.8064, and the root mean square error RMSE was 0.3251mg/g. The R2 of the validation set was 0.7915, and the RMSE was 0.3620mg/g. Therefore, the proposed AHA-ELM model established by DWT-DE transform had significant advantages in the rapid detection of rice nitrogen content, and can provide a good reference for precise variable fertilization in rice.

Key words:rice;nitrogen content;hyperspectral;coupled discrete wavelet transform and first-order differential transform;artificial hummingbird optimization;extreme learning machine