Recognition Of Wheat Diseases Based On Imagery And Spectral Analysis

The crop diseases are natural enemies of agriculture, restricting the grain production all the time. How to diagnose of disease severity scientifically and distinguish the similar diseases accurately is becoming an important agricultural application of remote sensing technology. The hyperspectral remote sensing can obtain both image information and spectral information of target, playing an important role in the diagnosis mechanism of crop disease and promoting the development of agricultural remote sensing technology. In this paper, the powdery mildew and yellow rust of wheat were selected as the object to study. Then the recognition model and distinguishing model at leave level and canopy level were set up respectively. The results provided references for portable instrument development and disease recognition in the early stage. The main results are as follows:1) 3 nm imaging hyperspectral data was used to detect the spectral features of wheat leaves affected with diseases. The image features with the yellow rust was studied, for example. The data ranges of the occurrence measures in window of 3×3 and the dissimilarity of the co-occurrence measures in window of 3×3 both had the best results, the details could reflect the most prominent characteristics. They could reflect the characteristics of the structure of the diseased area preferably, in order to analyze images of yellow rust and powdery mildew subsequently. The 110 pixels of the healthy leaves, the diseased leaves with slight yellow rust, the diseased leaves with severe yellow rust, the diseased leaves with slight powdery mildew and the diseased leaves with severe powdery mildew were selected respectively. On the basis of 10 near-infrared bands (706.2-712.1 nm) and 10 red bands (675.1-681.1 nm) which were sensitive to this disease, we built an image feature space (X axis:red; Y axis:NIR) to identify disease spots. The healthy pixels and the pixels of the slight diseased leaves had a little overlap.2) At the leave level, the dimensionality reduction of the hyperspectral images was done by using the principal component analysis (PCA). The recognition of the diseased area at leaf level was done by using the density slice method. On this basis, the spectral difference of two diseases was analyzed, and twelve of the disease-sensitive bands were selected in the light of the second principal component (PC-2) images. The bands of the powdery mildew were at 519,643,696,764,795 and 813 nm, while those of the yellow rust were at 494,630,637,698,755 and 805 nm. Furthermore, a discriminant model of the support vector machine (SVM) was established based on the sensitive wavebands, and its accuracy reached 92%. The discriminant model distinguished powdery mildew and yellow rust at leaf level effectively, which will provide a data support for developing a portable recognition device.3) At the canopy level, a hyperspectral imager SOC710-VP in this study was used to detect the extraction of wheat powdery mildew considering field comprehensive effects. Through comparing the spectral differences among healthy-, shadow-, infected leaves and wheat ear,23 sensitive bands were chosen to distinguish different target objects. And five kinds of vegetation indices (VIs) and three red edge parameters were calculated based on sensitive bands. And then 40 identification features were chosen to determinate different background targets and disease severity. Moreover, Classification and Regression Tree (CART) was utilized to construct estimation model of powdery mildew, the identification accuracy was assessed by cross-validation. These results showed that shadow leaf could be 100 percent recognized, while the discrimination accuracies are respectively 98.4%,98.4%,80.8% for healthy leaf, infected leaf and wheat ear. For different disease severity, healthy leaf has the highest estimation accuracy (99.2%), while moderate and mild infections are 88.8% and 87.9%, respectively. Thus, wheat ear is an important background factor to effect the detection of wheat powdery mildew using in-situ remote sensing technology. Meanwhile, how to improve the assessment accuracy for mild disease is also to pay attention.