基于小波特征的小麦白粉病与条锈病的定量识别

小麦白粉病和条锈病是小麦常发病害中为害较重的两种病害，在我国小麦产区均有发生，但它们由不同病原引起，需要采取不同的防治措施。因此，快速、准确的获取小麦病害类型信息对于病害的防治具有重要的指导意义。遥感数据具有快速、准确的获取空间上连续信息的特点，提出一种基于实测冠层高光谱数据信息的小麦病害定量识别方法。通过对标准化光谱进行连续小波变换，分析350～1 300 nm范围内各波段及其连续小波特征与小麦白粉病和条锈病之间的相关性，以及在不同病害间的差异性，筛选出对不同病害敏感的光谱波段(SBs)和小波特征(WFs)，然后采用Fisher判别分析法分别基于SBs，WFs以及结合SBs和WFs建立小麦白粉病、条锈病及正常小麦识别模型，分别采用未参与建模的55个地面调查数据和留一法进行验证。结果显示： (1)基于WFs模型的总体识别精度(分别为92.7%和90.4%)明显高于基于SBs模型的总体识别精度(分别为65.5%和61.5%);(2)SBs和WFs结合模型的总体识别精度(分别为94.6%和91.1%)略高于基于WFs模型的总体识别精度，在Fisher80-55模型中白粉病和正常样本的生产者精度提高了10%以上。(3)条锈病样本能在基于WFs和SBs & WFs的模型中准确判别出来，用户精度和生产者精度均达到100%。结果表明采用作物光谱信息能够准确的识别健康作物和不同类型的作物病害，为采用遥感影像进行大范围作物病害识别提供了理论基础，对于指导作物病害防治具有实际应用价值。

Quantitative Identification of Yellow Rust and Powdery Mildew in Winter Wheat Based on Wavelet Feature

Powdery mildew (Blumeria graminis) and stripe rust (Puccinia striiformis f .sp .Tritici) are two of the most prevalent and serious winter wheat diseases in the field ,which caused heavy yield loss of winter wheat all over the world .It is necessary to quantitatively identify different diseases for spraying specific fungicides .This study examined the potential of quantitative dis‐tinction of powdery mildew and yellow rust by using hyperspectral data with continuous wavelet transform at canopy level .Spec‐tral normalization was processing prior to other data analysis ,given the differences of the groups in cultivars and soil environ‐ment .Then ,continuous wavelet features were extracted from normalized spectral bands using continuous wavelet transform . Correlation analysis and independent t‐test were used conjunctively to obtain sensitive spectral bands and continuous wavelet fea ‐tures of 350 ~ 1 300 nm ,and then ,principal component analysis was done to eliminate the redundancy of the spectral features . After that ,Fisher linear discriminant models of powdery mildew ,stripe rust and normal sample were built based on the principal components of SBs ,WFs ,and the combination of SBs & WFs ,respectively .Finally ,the methods of leave‐one‐out and 55 sam‐ples which have no share in model building were used to validate the models .The accuracies of classification were analyzed ,it was indicated that the overall accuracies with 92.7% and 90.4% of the models based on WFs ,were superior to those of SFs with 65.5% and 61.5% ;However ,the classification accuracies of Fisher 80‐55 were higher but no different than leave‐one‐out cross validation model ,which was possibly related to randomness of training samples selection . The overall accuracies with 94.6% and 91.1% of the models based on SBs & WFs were the highest ;The producer’ accuracies of powdery mildew and healthy samples based on SBs & WFs were improved more than 10% than those of WFs in Fisher 80‐55 .Focusing on the dis‐criminant accuracy of different disease ,yellow rust can be discriminated in the model based on both WFs and SBs & WFs with higher accuracy ;the user’ accuracy and producer’ accuracy were all up to 100% .The results show great potential of continuous wavelet features in discriminating different disease stresses ,and provide theoretical basis for crop disease identification in wide range using remote sensing image .