【摘要】 【目的】白粉病严重危害小麦生长及制约产量形成，确立实时监测小麦白粉病的多源数据融合方法，为精确防控及保证国家粮食安全提供技术支撑。【方法】在小麦开花和灌浆期，使用同时搭载多光谱仪和热成像仪的六旋翼无人机作为遥感数据获取平台，通过ENVI软件从小麦白粉病遥感影像中提取植被指数、纹理特征以及冠层温度信息，进而利用多元线性回归（MLR）、后向传播神经网络（BP）、随机森林（RF）、极限学习机（ELM）算法将植被指数（VIs）、纹理特征（TFs）和温度特征（T）进行结合，以构建小麦白粉病病情指数的监测模型。【结果】无论是单数据源建模，还是多数据源建模，随机森林（RF）的精度均高于其他模型；3种数据源中植被指数的RF模型（VIs-RF,R~2=0.667,RMSE=5.712,RPD=1.572）更适宜白粉病监测，其次是温度特征（T-RF,R~2=0.559,RMSE=6.563,RPD=1.430），而纹理特征（TFs-RF,R~2=0.495,RMSE=7.014,RPD=1.348）效果最差；多数据源协同建模间比较，RF协同植被指数和纹理特征的模型R~2为0.701(VIs&TFs-RF,R~2=0.701,RMSE=5.308,RPD=1.724)，仅比VIs-RF模型R~2提升5.101%,RMSE降低7.073%,RPD提高9.672%，而RF协同植被指数和温度特征模型（VIs&T-RF）以及协同3种数据源模型（VIs&TFs&T-RF）的精度分别为R~2=0.750,RMSE=4.704,RPD=1.912和R~2=0.820,RMSE=4.677,RPD=1.996，较VIs-RF模型R~2分别提升12.453%和23.181%,RMSE分别降低17.640%和18.113%,RPD分别提高21.667%和26.981%。同时对不同模型进行10折交叉验证，进一步证实了RF模型在多数据源融合建模中性能稳定，估算效果最好。【结论】采用多数据源协同建模能够提升小麦白粉病遥感监测精度，研究结果为实现大面积高精度遥感监测作物病害状况提供了思路与方法。更多还原

【Abstract】 【Objective】Wheat growth and yield can be seriously affected by powdery mildew. Establishing the multi-source data fusion method for real-time monitoring of powdery mildew of wheat could provide technical support for accurate prevention and control of diseases and guaranteeing national food security.【Method】During the wheat flowering and filling period, a six-rotor UAV equipped with multi-spectral sensor and thermal imager was used as a remote sensing data acquisition platform to obtain remote sensing images of different degrees of wheat powdery mildew. Then, vegetation index(VIs), texture feature(TFs) and temperature feature(T) were extracted from multi-spectral and thermal infrared images of different disease degrees on a low-altitude drone platform by ENVI software. Finally, the wheat powdery mildew disease index model were built by multiple linear regression(MLR), back propagation neural network(BP), random forest(RF) and extreme learning machine(ELM). 【Result】The precision of the RF model based on both single and multiple data sources was higher than that of the other models. Among the three data sources of the RF model, the vegetation indices(VIs-RF, R~2 = 0.667, RMSE=5.712, RPD=1.572) were the most suitable for powdery mildew monitoring, followed by the temperature feature(T-RF, R~2 = 0.559, RMSE=6.563, RPD=1.430) and texture features(TFs-RF, R~2 = 0.495, RMSE=7.014, RPD=1.348). When combining multiple data sources, a precision for the RF model combining vegetation indices and texture features(VIs & TFs-RF) of 0.701 could be obtained, which was 5.101% higher than that of the VIs-RF model, while RMSE was 7.073% lower and RPD was 9.672% higher, whereas the precision parameters of the RF model combining vegetation indices and the temperature feature(VIs & T-RF) were R~2 = 0.750, RMSE = 4.704, RPD = 1.912. For all three remote sensing data sources(VIs & TFs & T-RF), the following accuracies resulted: R~2 = 0.820, RMSE = 4.677, RPD=1.996. As compared to the VIs-RF model, R~2 improved by 12.453%, RMSE by 17.640% and RPD by 21.667% for the(VIs & T-RF) model, whereas for the three remote sensing sources, R~2 improved by 23.181%, RMSE by 18.113% and RPD by 26.981%. At the same time, 10 fold cross validation of different models was carried out, which further confirmed that RF model had stable performance and good estimation results in multi-data source fusion modeling.【Conclusion】 The precision of wheat powdery mildew monitoring could be improved by using multi-data-sources collaborative ML modeling. This research provided technical support for large-area and high-precision remote sensing of crop diseases.更多还原