Study On The Color Detection And Early Blight Disease Identification Of Eggplant Leaves Based On Hyperspectral Imaging

Precision agriculture is the new direction for modern agriculture, and it requires informatization, automation and sustainability with the development of science and technology. Thus, traditional agriculture technologies and experimental analysis methods cannot meet the request for modern agriculture. To develop disease detection and monitor instruments has become the most important task for China and other countries in the world. The change of color values on eggplant leaves stressed by early blight disease was studied by using Vis/NIR and NIR hyperspectral imaging techniques, respectively. Also, the spectral reflectance, texture features extracted from grey images and RGB/HSV/HLS images,color features of RGB/HSV/HLS images were used for classifying healthy and diseased samples based on Vis/NIR and NIR hyperspectral imaging techniques. The obtained results could provide information for the growth of eggplant crops, and they also support developing disease detection and monitor instruments. The main content for this project can be seen as follows:(1) The three color values (L*, a* and b*) of healthy and early blight disease infected eggplant leaves were studied by Vis/NIR and NIR hyperspectral imaging techniques. It can be found that the average value of L* for healthy leaves is higher than that for diseased samples. This is because the higher L* stands for lightness, and the leaves became darker when they were infected by early blight disease. It is the same with the color a*, positive a* demonstrates the sample is redder than standard, and negative means greener than standard, leaves became ungreen when they were infected disease, resulting in the average value of a* lower for healthy samples. In Vis/NIR spectral region, LS-SVM model performed the best with the R2p of 0.660 and RMSEP of 1.766 for L*, LS-SVM model performed the best with the R2p of 0.869 and RMSEP of 2.068 for a*, CA-BPNN model performed the best with the R2p of 0.903 and RMSEP of 2.172 for b*. In NIR spectral region, CA-BPNN model performed the best with the R2p of 0.618 and RMSEP of 2.404 for L*, Normalization-PLS model performed the best with the R2p of 0.861 and RMSEP of 2.114 for a*, CARS-LS-SVM model performed the best with the R2p of 0.795 and RMSEP of 3.190 for b*.(2) Spectral reflectance extracted from Vis/NIR and NIR hyperspectral images were used for establishing classification models, respectively. In Vis/NIR spectral region, all models obtained the classification results over 96.18% based on full spectral wavelengths. The classification results were between 66.24% and 100% based on CARS, over 91.03% based on RC except RC-SVM model (Trainging set:56.05%, Testing set:56.41%) and over 78.98% based on CA method except the CA-SVM model (Trainging set:55.41%, Testing set: 55.13%). In NIR spectral region, all models obtained the classification results over 95.54% based on full spectral wavelengths. The classification results were between 93.59% and 100% based on CARS, between 88.46% and 100% based on RC and over 92.31% based on CA method. All classification models performed well except RC-SVM and CA-SVM models in the Vis/NIR spectral region.(3) Texture features extracted from Vis/NIR and NIR hyperspectral images were used for detecting early blight disease on eggplant leaves. Several gray images at the effective wavelengths suggested by RC were selected from Vis/NIR and NIR hyperspectral images. Texture features were extracted from the grey images based on occurrence measures and co-occurrence measures, respectively. Then, different classification models were established based on these texture features. It can be found all models performed well with the classification results between 83.33% and 100% except SVM model (Testing set:55.13%) for Vis/NIR hyperspectral images. The results were between 61.54% and 100% except SVM model (Testing set:55.13%) for NIR hyperspectral images.(4) Texture features extracted from RGB, HSV and HLS images were used for establishing classification models. Vis/NIR hyperspectral images were converted into RGB, HSV and HLS images firstly, then different texture features based on occurrence measures and co-occurrence measures were extracted from RGB, HSV and HLS images, respectively. Based on RGB images, all models provided good performance with the classification over 92.31% except SVM model (Testing set:55.13%). For HSV images, the results were over 93.59% except SVM model and occurrence measures-based AdaBoost model. The HLS images obtained the results over than 88.46% except SVM model (Testing set:55.13%).(5) The color features for RGB, HSV and HLS images were used for detecting early blight diseased samples. Based on these features, different classification models were built. The results were between 92.99% and 100% for RGB images, between 85.99% and 100% for HSV images and between 69.43% and 100% for HLS images.This study aimed at detecting early blight disease on eggplant leaves using hyperspectral imaging technique. Spectral reflectance, texture features from gray images, texture features from RGB/HSV/HLS images and color features from RGB/HSV/HLS images were studied and used for building classification models. Also, the color values (L*, a* and b*) were predicted by using Vis/NIR and NIR hyperspectral imaging techniques. This technique has the potential to be used for real time, on-line and nondestructive detection of early blight disease on eggplant leaves. Also, it was significant for developing disease detection and monitor instruments.