对水稻稻瘟病病害程度的定量预测是精准防控的关键, 田间冠层尺度的研究可为高光谱传感器提供理论基础。 以受穗颈瘟胁迫的水稻为研究对象, 采用SVC HR768i型光谱辐射仪在大田中获取灌浆期两个不同时间段的水稻冠层光谱反射率, 以水稻发病株数百分比作为病害严重程度指标。 冠层光谱数据采用九点平滑预处理, 并重采样为1 nm间隔, 计算植被指数; 经过去包络线和一阶导数光谱变换, 提取高光谱特征参数。 分析不同时间段的光谱变换、 植被指数、 高光谱特征参数与病害程度的相关关系, 构建基于植被指数、 高光谱特征参数的穗颈瘟病害程度随机森林预测模型, 并对比分析两个单时期预测模型异同, 优选共用输入量, 构建出两时期混合数据的病害程度预测模型。 结果表明: （1）原始光谱曲线经去包络线处理可有效增强与病害程度相关的光谱信息, 近红外波段（960～1 050和1 150～1 280 nm）的相关系数在0.80以上; （2）高光谱特征参数与病害程度相关性分析中, 去包络线吸收谷参数相关系数高于其他参数, 吸收谷V3（910～1 100 nm）、 吸收谷V4（1 100～1 300 nm）中面积（A3和A4）、 深度（DP3和DP4）、 斜率（SL4和SR4）的相关系数在0.74以上; （3）去包络线吸收谷参数结合随机森林模型预测穗颈瘟病害程度在单时期及两时期混合数据中均表现最好。 灌浆期后期数据预测效果最佳, 验证集决定系数R2=0.91, 均方根误差RMSE=0.02; （4）两时期混合数据预测精度处于两个单时期预测精度之间, 验证集决定系数R2=0.85、 均方根误差RMSE=0.03。 研究成果揭示了灌浆期不同时间段水稻穗颈瘟光谱响应机制, 表明去包络线吸收谷参数结合随机森林模型预测稻瘟病的实用性, 可为田间水稻穗颈瘟病害程度进行快速、 精确、 无损地定量预测, 为精准施药提供理论依据, 并对未来航空、 航天遥感的病害监测提供一定的技术支持。

Quantitative prediction of disease degree of rice panicle and neck blast is essential on accurate prevention and control measures. The study of field canopy scale can provide a theoretical basis for hyperspectral sensors. In this paper, the rice which was damaged by panicle and neck blast was regarded as the research object, and hyperspectral canopy reflectance was acquired by SVC HR768i spectral radiometer at two different periods during the filling stage. The percentage of rice plants diseased represented disease degree index. The canopy spectral data were preprocessed by nine-point smoothing and resampled at 1 nm intervals. Vegetation indexes were calculated and hyperspectral characteristic parameters were extracted by continuum removal (CR) and first derivative reflectance. Were totally analyzed between each period, the response ability of different spectral transformation, vegetation index and hyperspectral characteristic parameters to disease degree through correlation analysis, and prediction models of disease degree were established through the random forest (RF) based on vegetation index and hyperspectral characteristic parameters, respectively. The two single-period prediction models were compared to select the common input to generate a disease degree prediction model which mixed data in two periods. The results demonstrated that: (1) Canopy hyperspectral reflectance processed by continuum removal (CR) method could effectively enhance the spectral information which isclosed related to the disease degree. The sensitive bands were the near-infrared region (960～1 050 nm) and (1 150～1 280 nm), and the correlation coefficient was above 0.80. (2) In the correlation analysis between hyperspectral characteristic parameters and the disease degree, the correlation coefficient of absorption valley parameters extracted by CR was higher than other parameters, and that of area (A3, A4), depth (DP3, DP4) and slope (SL4, SR4) in the absorption valley V3(910～1 100 nm) and V4(1 100～1 300 nm) was above 0.74. (3) The absorption valley parameters which played a role as the model input showed the best result in the mixed data of two periods and that of every single period. In addition, the prediction accuracy reached a peak at the later filling stage, with R2=0.91 and RMSE=0.02 in the validation set. (4) The prediction accuracy of the mixed data of two periods was between that of two single-period, with R2=0.85, and RMSE=0.03 in the validation set. The results revealed the spectral response mechanism of rice panicle and neck blast at different periods during the filling stage and it was practical to predict disease degree by combining absorption valley parameters extracted by CR with the random forest model, which can be used to rapidly, accurately and nondestructively predictthe disease degree of rice panicle and neck blast and provided a theoretical basis for precise application of pesticides. Beyond that, it also provided some technical reference for aviation and aerospace remote sensing monitoring in the future.