摘要
小麦受到条锈病菌侵染后,作物的光合能力及色素含量等均会发生变化,日光诱导叶绿素荧光(solar-inducedchlorophyll fluorescence,SIF)对作物光合生理的变化比较敏感,而反射率光谱则受作物生化参数的影响较大,为了提高小麦条锈病的遥感探测精度,该文利用随机森林(random forest,RF)等机器学习算法开展了协同冠层 SIF 和反射率微分光谱指数的小麦条锈病病情严重度的遥感探测研究。首先利用 3FLD(three bands fraunhofer line discrimination)算法提取了冠层SIF 数据,然后结合对小麦条锈病病情严重度敏感的 11 种反射率微分光谱指数分别基于 RF 和后向传播(back propagation,BP)神经网络算法构建了反射率微分光谱指数与冠层 SIF 协同的小麦条锈病病情严重度预测模型。研究结果表明:RF 算法构建的小麦条锈病病情严重度预测模型优于 BP 神经网络算法,3 个样本组中 RF 模型病情指数(disease index,DI)估测值与实测值间的决定系数 R2平均为 0.92,比 BP神经网络模型(R2的平均值为 0.83)提高了 11%,均方根误差(root meansquare error,RMSE)平均为 0.08,比同组 BP 神经网络模型(RMSE 的平均值为 0.12)减少了 33%,RF 算法更适合于小麦条锈病病情严重度的遥感探测。在反射率微分光谱指数中加入冠层 SIF 数据后,RF 模型和 BP 神经网络模型精度均有所改善,其中 RF 模型估测值与实测值间的平均 R2提高了 4%,平均 RMSE 减少了 22%,BP 神经网络模型估测值与实测值间的平均 R2提高了 14%,平均 RMSE 减少了 28%,综合利用冠层 SIF 和反射率微分光谱指数能够改善小麦条锈病病情严重度的遥感探测精度。研究结果可为进一步实现作物健康状况大面积高精度遥感监测提供新的思路。
The prevalence of wheat stripe rust has a significant impact on the production of winter wheat all over the world. An effective monitoring and warning of this disease is imperative to ensure the quality of wheat production. Remote sensing detection of wheat stripe rust is important for agriculture management and decision. The reflectance spectrum is closely related to the changes of biomass. It cannot, however, directly reveal the photosynthetic physiological state of vegetation. Solar-induced chlorophyll fluorescence(SIF) can sensitively reflect the photosynthetic vitality of crops, and the canopy’s solar-induced chlorophyll fluorescence signal includes the fluorescence characteristics of physiological changes caused by plant disease stress. In order to improve detection precision of wheat stripe rust, this study made full use of the advantages of reflectance spectroscopy for the detection of crop biochemical parameters and the advantages of chlorophyll fluorescence in photosynthetic physiological diagnosis, a remote sensing study on the severity of wheat stripe rust was carried out by using random forest (RF) and other machine learning algorithms synergistic SIF and reflectance differential spectral index in the canopy of wheat. Firstly, based on Fraunhofer line principle, three bands fraunhofer line discrimination(3FLD) algorithm was used to predict the intensity of chlorophyll fluorescence in O2-A band (760 nm). Then 11 reflectance differential spectral indices, which are sensitive to the severity of wheat stripe rust disease were selected. Based on RF and back propagation(BP) neural network algorithm, a model for predicting the severity of wheat stripe rust with differential reflectance spectral index and canopy SIF was established. The study incorporated a cross-checking method based on measurements of control samples. Fifty-two raw crop samples were randomly divided into two parts three times, the first part including 39 datasets was used as the training set for the model building, and the remaining 13 data samples were used to evaluate the accuracy of the models. The results showed that: 1) There is a significant negative correlation between SIF and the disease severity of wheat stripe rust. Remote sensing detection of wheat stripe rust severity can both be realized using the differential spectral index alone or by using the differential spectral index and the solar-induced chlorophyll fluorescence in combination. However, the accuracy of the estimates made by the RF and BP neural network models using the combination of data from the differential spectral index and the solar-induced chlorophyll fluorescence were all higher than that for the models constructed using the differential spectral index alone. In the three sample groups, average determination coefficient between the estimated DI using the RF model and the BP neural network model and the measured DI increased by 4% and 14% respectively, and the average RMSE decreased by 33% and 28% respectively. The detection accuracy of wheat stripe rust severity can be improved using solar-induced chlorophyll fluorescence combined reflectance differential spectral index. 2) The canopy solar-induced chlorophyll fluorescence synergistic differential spectral index were used as sensitive factors, the coefficients of determination between the estimated DI using the RF model and the measured DI were 0.90, 0.93, and 0.98, respectively, which were greater than the coefficients produced when using the BP neural network model for the same group (0.88, 0.84, and 0.92). Similarly, the RMSEs were 0.09, 0.07, and 0.04, respectively, which were smaller than the RMSEs (0.10, 0.11, and 0.09) using the BP neural network model for the same group. Therefore, the model using the RF algorithm was better at estimating wheat stripe rust severity than the BP neural network-based model, and it is more suitable for the remote sensing detection of wheat stripe rust severity. These results have important significance for improving the accuracy of the real-world remote sensing detection of wheat stripe rust, and the analysis provides new ideas for further realizing large-area remote sensing monitoring of crop health.
作者
竞霞
白宗璠
高媛
刘良云
Jing Xia;Bai Zongfan;Gao Yuan;Liu Liangyun(College of Geometrics, Xi’an University of Science and Technoology, Xi’an 710054, China;Institute of Remote Sensing and Digital Earth, Chinese Academy of Science, Beijing 100094, China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2019年第13期154-161,共8页
Transactions of the Chinese Society of Agricultural Engineering
基金
国家自然科学基金（41601467）
关键词
遥感
监测
小麦
条锈病
随机森林
日光诱导叶绿素荧光
微分光谱指数
remote sensing
monitoring
wheat
stripe rust
random forest
solar-induced chlorophyll fluorescence
differential spectral index
