基于相关向量机的冬小麦蚜虫遥感预测

Forecasting wheat aphid with remote sensing based on relevance vector machine

蚜虫的流行严重影响了冬小麦的产量。区域尺度上及时准确的预报冬小麦蚜害发生范围能为蚜害的有效预防提供基础信息,从而降低冬小麦产量的损失。该研究利用多时相的环境星CCD光学数据和IRS热红外数据,分别提取了冬小麦的长势因子,比值植被指数(ratio vegetation index,RVI)和归一化植被指数(normalized difference vegetation index,NDVI),以及生境因子,地表温度(land surface temperature,LST)和垂直干旱指数(perpendicular drought index,PDI),利用相关向量机(relevance vector machine,RVM)、支持向量机(support vector machine,SVM)和逻辑回归(logistic regression,LR)方法建立了北京郊区冬小麦灌浆期蚜虫发生预测模型,并对比分析了3种模型预测精度。试验结果表明,RVM总体预测精度达到87.5%,优于SVM的79.2%和LR的75.0%。另外,RVM模型计算量较小,相比于SVM和LR模型,其预测时间可分别缩短7倍和60倍。较高预测精度和较小计算量的特性扩大了RVM在实际中的应用价值。

The prevalence of aphid in winter wheat field has a significant impact on the production of winter wheat. An effective and timely forewarning of the scope and severity of the disease at a regional scale will not only reduce yield losses but also alert the stakeholders to take effective preventive measures. Forecasting aphid occurrence with Remote Sensing has great advantages over traditional methods and meteorological data, such as lower cost, simpler operation, more real-time and higher resolution. Chinese HJ-1A/1B data which have a high revisit frequency （〈4 days ）as well as 30 m spatial resolution, in addition to a full band set covering visible, near-infrared, short-wave infrared, mid-infrared and thermal infrared spectral ranges, which thus significantly expand opportunities of remote sensing application in predicting crop diseases and pest. According to the literature research and field investigation found that apart from crops growth status was closely related to wheat aphid occurrence, development, and dispersal, the environmental conditions of field, such as temperature and humidity ,also have a certain impact on its susceptibility to pest. In this study, we retrieve two parts of information from multi-temporal HJ-CCD optical data and HJ-IRS thermal infrared data, including Ratio vegetation index （RVI） and Normalized Difference Vegetation Index （NDVI） representing grow status of wheat, Perpendicular Drought Index （PDI） and Land Surface Temperature （LST） that represent the environmental characteristics. An independent t-test analysis was used to test the difference between pest and healthy samples based on calibration data. Those vegetation and environmental factors that failed to show a statistical significant （p-value 〈0.001） were eliminated. Five remotely sensed variables （NDVI-T1, RVI-T1, LST-T1, LST-T2, and PDI-T2） were identified as optimal explanatory variables for developing the pest forecasting model. The relevance vector machine （RVM） model was established to predict aphid occurrence of filling stage of wheat in the Beijing suburbs, which is a machine learning algorithm and commonly used to improve business decisions, detect disease, and forecast weather credit to its superior learning ability when on a small datasets. The results obtained from the RVM model are compared with prediction models developed using logistic regression （LR） and support vector machine （SVM） techniques. Goodness-of-fit values of the calibrated models were evaluated through the p-value of Spearman test. In addition, the association between observed outcome of pest occurrence and predicted outcome of pest occurrence was evaluated by using three different measures： Somers＇D, Goodman-Kruskal gamma, and Kendall＇s Tau-c. The RVM model produced a higher Spearman p-value than that the LR model and the SVM model did, indicating that the RVM model had a better performance to predict the aphid than the other two models. Moreover, the values of Somers＇D, Goodman-Kruskal Gamma, and Kendall＇s Tau-c were all higher with the RVM model than those with the remaining model, which further demonstrated the better performance of the RVM model on the training datasets. To further evaluate the difference of performances of the three types of models, we obtain accuracy of models respectively based on validation samples. The results showed that： accuracy by using RVM algorithm is the highest among the three methods, with lower omission and commission than the other two algorithms. Furthermore, the overall prediction accuracy and the kappa coefficient of the RVM model are 87.5%, 0.71respectively, have shown better performance over the SVM model （79.2%,0.55） and the LR model （75.0%,0.44）. Additionally, the RVM model needs less computation work and faster prediction speed. These results revealed that the RVM model with high accuracy level and prediction efficiency is much more favorable in forecast the pest occurrence at a regional scale effective and timely. It indicated that the result of this research can be used in different regions.