Quantitative identification of crop disease and nitrogen-water stress in winter wheat using continuous wavelet analysis

It is necessary to quantitatively identify different diseases and nitrogen-water stress for the guidance in spraying specific fungicides and fertilizer applications. The winter wheat diseases, in combination with nitrogen-water stress, are therefore common causes of yield loss in winter wheat in China. Powdery mildew (Blumeria graminis) and stripe rust (Puccinia striiformis f. sp. Tritici) are two of the most prevalent winter wheat diseases in China. This study investigated the potential of continuous wavelet analysis to identify the powdery mildew, stripe rust and nitrogen-water stress using canopy hyperspectral data. The spectral normalization process was applied prior to the analysis. Independent t-tests were used to determine the sensitivity of the spectral bands and vegetation index. In order to reduce the number of wavelet regions, correlation analysis and the independent t-test were used in conjunction to select the features of greatest importance. Based on the selected spectral bands, vegetation indices and wavelet features, the discriminate models were established using Fisher's linear discrimination analysis (FLDA) and support vector machine (SVM). The results indicated that wavelet features were superior to spectral bands and vegetation indices in classifying different stresses, with overall accuracies of 0.91, 0.72, and 0.72 respectively for powdery mildew, stripe rust and nitrogen-water by using FLDA, and 0.79, 0.67 and 0.65 respectively by using SVM. FLDA was more suitable for differentiating stresses in winter wheat, with respective accuracies of 78.1%, 95.6% and 95.7% for powdery mildew, stripe rust, and nitrogen-water stress. Further analysis was performed whereby the wavelet features were then split into high-scale and low-scale feature subsets for identification. The accuracies of high-scale and low-scale features with an overall accuracy (OA) of 0.61 and 0.73 respectively were lower than those of all wavelet features with an OA of 0.88. The detection of the severity of stripe rust using this method showed an enhanced reliability (R-2 = 0.828).