基于高光谱成像的寒地水稻叶瘟病与缺氮识别

为进行水稻叶瘟病与养分缺失的区分、实现叶瘟病及时、准确的诊断,以大田试验为基础,利用高光谱成像仪获取2个品种的健康、缺氮、轻度感病和重度感病共4类水稻叶片的反射率光谱,对其光谱特性进行分析,并采用多种预处理方法、分别结合偏最小二乘判别分析(partial least squares-discriminate analysis,PLS-DA)和主成分加支持向量机(principle component analysis-support vector machine,PCA-SVM)方法构建水稻叶瘟病识别模型。试验结果显示6个判别模型都获得了较高的识别准确率,经标准正态变量(standard normal variate,SNV)变换预处理的PLS-DA模型获得了最佳的识别结果,预测准确率达100%,经多元散射校正(multiplicative scatter correction,MSC)预处理的PCA-SVM模型的预测准确率也达到97.5%。本研究为水稻叶瘟病的判别和分级提供了新方法,也为稻瘟病大范围遥感监测提供了基础。

Identification of rice leaf blast and nitrogen deficiency in cold region using hyperspectral imaging

Rice blast is one of the most serious rice diseases and significantly impacts rice yields. In recent years, it is a hotspot to use hyperspectral imaging technology for the non-destructive identification of rice blast. However, nutrient deficiency in rice(such as nitrogen, potassium, etc.) will probably result in the chlorosis similar to rice blast. Therefore, to differentiate between them is very important for field management. In this study, field trials of rice blast and nitrogen stress were carried out in Fangzheng, Harbin, and 2 rice varieties with weak resistance were involved. From 8 to 10 in July, 2015, 4 types of rice leaves from both 2 varieties, including 60 in group of health, 60 in group of nitrogen deficiency, 60 in group of mild infection and 60 in group of severe infection, were collected and their hyperspectral images were captured with the HeadWall hyperspectral imaging system, and then the average reflectance spectrum of interest region of different leaves were acquired using the environment for visualizing images. In order to explore 4 types of spectral characteristics, the average spectrum of each type sample data, which was smoothed with polynomial convolution smoothing(Savitzky-Golay, SG), were calculated as a spectral curve of each category. Significant differences were found at the following three positions: the range around 560 nm in the reflection peak of green wavelength region; the range from 620 nm to 670 nm in red wavelength region; and particularly remarkable in the range around 760 nm in high reflectance of the near infrared region. The models of rice leaf blast recognition were established by taking advantage of a partial least squares-discriminate analysis method(PLS-DA) and the principle component analysis plus support vector machine(PCA-SVM), and using three different data pretreatment methods to preprocess original reflectance spectrum data, i.e., SG, standard normal variate transformation(SNV) and multiplicative scatter correction(MSC). The models were tested with the cross-validation strategy. The key of PLS-DA model is to select appropriate number of factors, 20 of which were determined by repeated testing, and the PLS-DA models were established by Fisher method. The prediction effects of the three models with preprocessing spectrum were all greater than 96.3%, and better than the original reflectance spectra. The PLS-DA model with SNV pretreatment got the best discrimination results and the prediction accuracy rate was 100%. The PCA-SVM models were constructed by taking the first 15 principal components as inputs and by adopting radial basis function as kernel function with the penalty coefficient was 10000 and kernel radius was 0.00599. The prediction accuracy of the three models with preprocessing spectrum were all equal to or greater than 95%, also better than the original reflectance spectra, and the discrimination results of PCA-SVM model with MSC pretreatment was 97.5%. This study provides a new idea and method for the nondestructive detection and identification of rice leaf blast, and also lays a foundation for a wide range of remote sensing and monitoring for rice blast.