Abstract:
Objective A nondestructive, effective method was developed based on the hyperspectral imaging technology for early diagnosis of the highly destructive wet bubble disease on Agaricus Bisporus caused by Mycogone perniciosa.
Method Information on the full band (401-1046nm) visible/near-infrared hyperspectral images on 200 healthy and 200 infected A. bisporus specimens was collected. After a preprocess using Savitzky-Golay 1st order derivative, Savitzky-Golay smoothing, or multiple scattering correction (MSC) on the obtained information of the 360 full bands, accuracy of the methodology in separating the healthy from the infected samples was scrutinized by using the Random Forest (RF), Support Vector Machine (SVM), and Extreme Learning Machine (ELM) models.
Result The 3 models yielded similar results and the MSC-SVM combination had the best detection effect with the identification accuracy on the test set increased from 85.02% to 92.21%, and on the prediction set, from 87.38% to 91.04%.
Conclusion The MSC-SVM model appeared to significantly improve the identification accuracy using the full band. It provided a basis for the development of rapid, nondestructive diagnostic device on the devastating disease of A. bisporus at early stage which has been conventionally conducted by expert visual examination, PCR analysis on the internal transcribed spacer gene, or traditional Koch's postulation.
下载: