|M.Sc Student||Ya'ackoby-Hammami Hadar|
|Subject||Detection of Diseases in Citrus and Mango fruits by|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Raphael Linker|
|Full Thesis text - in Hebrew|
In this thesis we investigated the use of hyperspectral imaging in the range of 400-720nm for detecting diseases on fruits. In order to minimize the "flux density effect" due to variation of the illumination angle because of the fruit curvature, we used the co-called R2α transformation, which is based on the ratio of wavelet coefficients at different levels.
We conducted as first experiment with citrus inoculated with mycelium of Penicillium digitatum fungus and a second experiment with mango inoculated with Alternaria Alternata. In both cases the fruits were inoculated by puncturing the fruit peel and applying on the puncture ?10l of suspension containing 106 spores per ml in citrus, and 105 spores per ml in mango. The inoculated fruits were stored in wet environment at 20ºC and hyperspectral images of the fruits were recorded on three different days.
The citrus images were analyzed according to individual wavelengths and the classification was done after R2α correction at five wavelengths 500,510,520,530,540 nm. Classification was performed using the Uniformly Most Powerful Test approach (test of known power) using selected pixels of intact and diseased regions of a fruit for calibration. The best results were obtained using the 540 nm wavelength, for which we observed an increase of the number of pixels recognized as "disease" as the experiment progressed. For the control fruit, the percentage of "disease" pixels remained low (~1%).
The mango images were first analyzed without applying the R2α transformation, and as expected the results were very poor. The classification was then done after applying the R2α transformation and using individual wavelengths (480, 490 and 500 nm). Good classification results were obtained, with the number of disease pixels increasing with time in inoculated fruits while the number of disease pixels in the control fruits remained small. Combining the R2α coefficients of two or three wavelengths using a linear classifier did not improve the classification. Using a neural network to combine the information of the R2α coefficients led to very good resutls, with a clear increase of the pixels recognized as "disease" in the infected fruits, while for the control fruit there was almost no change in the percentage of "disease" pixels.
This research shows that hyperspectral imaging can be used to identify disease in fruits. However, for on-line classification one should improve the algorithm, especially the R2α transformation which requires lengthy computations but is necessary because of the fruit curvature.