基于BP神经网络的青稞白粉病早期感染高光谱遥感监测

Identifying early infection of powdery mildew of highland barley based on BP neural network and hyperspectral remote sensing

关键词:  青稞  白粉病  神经网络  作物病虫害遥感

Key words:  highland barley  powdery mildew  neural network  remote sensing for crop diseases and pests

摘要:

【目的】 文章旨在探索表征西藏青稞白粉病早期感染的光谱特征参量，尝试基于机器学习算法构建白粉病高光谱遥感诊断模型的可行性，以期为经济、高效、及时、大区域的青稞白粉病早期遥感监测探索技术方法。【方法】 以西藏青稞主栽区和白粉病高发区拉萨、林芝和山南市为研究区域，开展青稞冠层光谱、病情和农学参数田间观测。提取冠层反射率（R）、反射率一阶导数（R′）、5种植被指数（VI），并联合应用R（或R′）与VI，共形成5套备选光谱特征参量。利用Relief-F算法开展特征选择，遴选白粉病感染的敏感波段和敏感特征。在此基础上，基于BP神经网络，构建白粉病诊断模型。【结果】 （1）对于染病样本而言，基于R′ 的诊断模型识别精度最高（51.33%），其次为R′与VI相结合的诊断模型（49.00%）。（2）R′ 的敏感波段范围为820~920 nm、1 160~1 200 nm和935~975 nm，此外445~490 nm和1 070~1 100 nm的R′也对白粉病表现出一定的敏感性。（3）5种VI中，比值植被指数（SR）对白粉病早期感染最为敏感，远超其它植被指数。此外，基于不同特征参量的健康样本的识别精度大致相当，在91.67%（基于R）~90.63%（基于R与VI）之间。【结论】 基于敏感波段反射率的一阶导数，利用构建诊断模型，有望有效地识别青稞白粉病早期感染。未来研究中，应考虑高光谱数据不同波段之间的高度相关性，开展数据降维，以期减小过拟合，进一步提高监测精度。

Abstract:

［Purpose］ This study aimes to explore spectral characteristic parameters that sensitive to early infections of powdery mildew of highland barley，and test the feasibility of identifying the infections using BP neural network and hyperspectral remote sensing techniques，in order to develop inexpensive and effective techniques to identify early infections of powdery mildew of highland barley in a large scale.［Method］ The study area was located in the the main producing area of highland barley and the high incidence area of powdery mildew in Tibet Autonomous Region，China，including Lhasa，Linzhi and Shannan City. In-situ measurement of canopy spectra，disease，and some agronomic parameters were performed in sample plots in the aforementioned study area. 5 sets of potential characteristic parameters，including canopy reflectance and its first derivative，5 vegetation indices（VIs），and the combination of them were tested. The method of Relief-F was used to select characteristic parameters that sensitive to the infection，which was then used as inputs of back propagation（BP）neural network to construct diagnostic model of powdery mildew infection.［Result］ To the infected samples，the first derivative of canopy spectra（R′）yielded the optimum accuracy of identification（51.33%），followed by the combination of R′ and VIs（49%）. The sensitive bands of R′ were 820~920 nm，1 160~1 200 nm and 935~975 nm. Besides，445~490 nm and 1 070~1 100 nm of R′ also showed certain sensitivity to powdery mildew. Among 5 VIs，simple ratio（SR）was the most sensitive to the infection，far beyond other VIs. Morever，different characteristic parameters yielded generally equivalent accuracy of healthy samples identification，ranking from 91.67%（based on canopy spectra，R）to 90. 63%（based on the combination of R and VIs）.［Conclusion］ According to this study，it is feasible to identify the early infection of powdery mildew of highland barley based on BP neural network and R′ of sensitive bands. In further studies，considering the high corelation between hyperspectral bands，it is considered helpful to reduce the dimensionalities of characteristic parameters，in order to reduce the risk of overfitting，and further improve the accuracy of identification.