 Thank you for the opportunity to present this research. I'm Nitin Nitin, Professor of the Department of Food Science and Technology and Bioengineering at University of California, Davis. And I'll be sharing our progress that we made in the area of spectroscopy and data analytics for the early detection of red blood viral infections in grid points. So to start with acknowledgement, I'd like to acknowledge the efforts of Dr. Reva Rai who is a project scientist at UC Davis and Christian. areas who is a post doctor scholar and my collaborators, my source to Darshan and Anita of a hospital at UC Davis and USDA RS. And also the funding from the SCRI program at USDA. Now, one of the key challenges in detection of grapevine red blood viruses is that these viruses have significant persistence and prevalence in many areas of growing regions. They can reduce the yield to the berries significantly also reduce the berries sugar content and then to sign in content to the skin. One of the challenges there is that these viruses have very similar features to other viruses such as deep fold disease. And these features are often visible only in the late season. And that is early fall of the crop. In addition to that, there are other nutrient stresses and other factors, agronomical factors that can produce similar kind of redness or blotch kind of features. So it makes it difficult to detect these disease visually among the wines. Now the current methods for detection has been based on real time PCR or PCR combined with electrophoresis. For these detection method these required verified DNA from plant samples to be isolated followed by RT PCR or PCR with molecular grade reagents and has to be done in a lab environment. There have been efforts to make these PCR based detection methods more feel deployable. For example, this is one of the example where the researchers are using the plasmon effect between the gold nanoparticles and the CRISPR gas to enable detection of the virus using visual detection, which requires isolation of DNA and their amplification could be done by PCR or other isothermal amplification processes. So our efforts have been devoted to evaluating if spectroscopy and imaging approaches can enable detection. The goal in spectroscopy approaches is to measure the spectral features of the between the control and the diseased leaves and identify if those spectral features can be used for detection. For example, shown here is a plot of IR spectroscopy, which is which could be collected by both lab based or the handheld devices. For spectroscopy we use mid IR wavelength, ranging from 2500 to 25,000 nanometer and also related to that is also visible and near IR spectroscopy that covers from 400 to 1100 nanometer in range. For spectroscopy you can collect imaging data, a simplest form of imaging would be an RGB image where you collect the information in red, green and blue channels and a more advanced feature would be collecting a hyper spectral image where you collect the spectral band across multiple wavelength regions and some of the hyper spectral imaging cameras can range from 400 nanometer to 2500 nanometers and in the visible to near IR regions, some of them are in the range of 400, 2000 nanometer regions. So to start with I'll talk about in detail about our efforts in the IR spectroscopy for red plot detection using FTIR and also discuss the potential of NIR spectroscopy. Now, just to give a primer on how IR spectroscopy works and how it could enable us to detect red blotch infections. So, typically in this case we send an IR source light from the source, select my point pointer. So we select a source of IR radiation that as it transmits through it excites the vibrational and the body energy of these molecules and some of the part of the spectrum are absorbed. So we measure the absorption of these molecules due to vibration of the bonds in this area. And now, as I mentioned we could do FTIR which is the mid IR range and also we could look at the visible to NIR region. So in terms of sensitivity FTIR has a much higher resolution into defined band with very strong absorption in the visible to near IR region the sensitivity is limited you have a broad bonds bands. And the functional groups that you can image with FTIR range from large diversity of functional group from carbonyl to hydroxyl to NH to CN and other groups. While the visible to near IR is limited to OH, CH or NH bonds, but they are also overlapping regions because they're not distinct bands sometimes in this case. The penetration depth of the FTIR is limited due to very high absorption and also the absorption of water and does require sample preparation so sample has to be isolated and dried. While the visible and to near IR the penetration depth can be more than three millimeter again depends on the object that you're looking at and it's more field deployable so limited sample preparation that sense. So we're starting with the mid IR region FTIR. What we did in this study was we took the grape leaves and we extracted the water extract after homogenization process and look at the composition analysis or the water extract using IR spectroscopy. Now this could be done in a lab environment using Benchtop IR meter or could be done in a handheld IR device in a field setting where the sample may be collected and tried using a small equipment in the field. The characteristic bands are shown here for the control and the GRBB infected leaves. So again the spectrum range from in wave numbers from 500 all the way to 4000 centimeter inverse. Now, we have done a multi-year field studies looking at the potential of this IR spectroscopy using the water extract of the leaves to enable detection of this. The samples have been collected from two different regions, one in the fertile region other is a NAPA region. The difference is being is that the fertile region is neither irrigated nor grafted while the NAPA region the wines are irrigated and grafted. We have collected over a period of time from May to October in both cases. The early stages that I refer to here refer to May and June and the late stages are September to October in these cases. So, if we take the FTIR spectrum and to distinguish between the control and the infected sample that is a non-infected and infected we take a second derivatives as a function of wave numbers of the spectrum. Here are shown are the differences in the second derivatives the spectral features of the infected versus non-infected wines in this case. And what we can find out is that there are significant differences in the in the bands, for example in the polysaccharide region in this shown here, as well as in the monoseccharide and the some of the polyphenols in this region of the spectrum. And these are detected again in the early season changes that we can detect between the wines when there are no visible symptoms on the wines in this case. And now here is a second derivative as a function of wave numbers for an irrigated wine that is collected in the NAPA region. Similarly, there are differences observed in the carbohydrate region as well as in the fingerprint region with monoseccharides and some of the polyphenols in this section of the spectrum. So clearly there is a potential of to detect these changes between the infected and non-infected wines based on the spectral features. Now if you were to follow this with the further analysis of PCA of spectral data, what you find is that we can see the differences between the non-infected indicated in blue versus the infected which are indicated in red in this case. Similarly, both in the non-irrigated and irrigated wines we were able to see differences between non-infected and the infected wines based on the spectral data and their PCA analysis. And in the non-irrigated wine it's fully emphasized that PC1 single component explained 91% of the differences between them, while in the irrigated wine it was more than 60% of the difference was explained by the single component in this case. So if you were to further analyze this using hierarchical clustering of infected versus non-infected wines in the early season, what you can clearly see is that the spectral differences can be used to discriminate the non-infected from the infected, both for the irrigated samples as well as a non-irrigated samples in the early season. As the time progresses, now I'm looking at this data in June, we can further see that the difference between the infected and the non-infected on the PCA plot increases significantly as also the discrimination power of the hierarchical clustering as shown here in this scenario. And again the PC1 is able to predict over 82% of the features in this case for the PCA analysis. Now this is in the late season of the irrigated wine again we can discriminate the non-infected which are in blue versus the infected and using hierarchical clustering again this further emphasizes that it's relatively easier to discriminate between the non-infected and infected wines using the spectral measurements using IR spectroscopy. And as the time progresses, the discrimination power of the analysis increases as you can see the clustering of the non-infected is limited to certain sections while the infected are also clustered more closely together and significantly far away from the non-infected wines. One of the potential approaches is to automate this analysis and to develop machine learning models in this direction we have looked at supervised learning for detection where we take the second derivative data and we have done the train and test set split between 70 to 30 using random stratified splits and then over 100 trials and 5-4 cross validation to evaluate the results. I'm going to share briefly the results of these prediction using decision tree and random forest, we're able to achieve very high degree of accuracy over 90-95% where we are able to discriminate the infected from the non-infected wines both with decision tree models and the random forest models further highlighting the ability of these models to predict the red blotch infection based on the spectral data. We have also taken the concept of near IR spectroscopy because as I indicated earlier these spectral measurements are more portable and can be done on an intact leaf without extracting the sample without drying. So we have made those measurements using portable near IR spectroscopy here are the results for the late season and what we see there is that we don't see significant differences in the near IR region. However, since we are looking at the visible to near IR which starts from 400 we do see some start seeing some changes in the visible region between the non-infected and the infected wines, great samples. And on the PCA plot we can clearly see the clustering of the non-infected which are again in blue and the clustering of the positive samples significantly distinct from each other. We were not able to see similar features in the early stages of the diseases. So when the early stages of the infection the visible to near IR spectroscopy were not able to detect these differences, possibly these differences point to the presence changes in the pigmentation of the leaves that may be becoming more visible during the late stages or the late season process. Moving forward, I'd like to talk about the hyperspectral and RGB imaging where we are using the imaging approach complementary to spectroscopy approach to enable the detection of red blotch infection in great points. Now just to give a brief introduction to RGB and hyperspectral imaging, in a simple RGB camera you are collecting an image of an object and dividing that image on the image plane into three different channels, the red, green and blue. In a hyperspectral imaging it's a little bit more involved in a sense that you collect the image then you use a dispersive element to break that into with the help of series of band pass filters to help that break into images at different spectral wavelengths. And as you can see, we are in the RGB you have three channels red, blue, green and red. Here you have multiple channels that you collect from the visible to the near IR region and this can go all the way to 1000 nanometer or to cover the full and IR range this can go all the way to 2,500 nanometer. Now thinking about how these compare to each other. In terms of penetration that if you're looking at are you collecting signals from the plant surface but also some some signal from the depth of it hyperspectral imaging using visible to me IR can allow us to deep penetrate light deeper as compared to the visible region. Now, in terms of functional properties pigments and diverse chemical features can be imaged by hyperspectral imaging, while in RGB channel we are limited to predominantly pigment change in changes but also some chemical features could be detected in that region also. In terms of cost hyperspectral imaging systems are relatively high cost, while the RGB imaging systems are relatively low cost and could be just as simple as your cell phone camera with the modified filters. Now to illustrate how does the hyperspectral imaging of leaf look like. Now here I'm showing the images of infected and non infected leaf late in the season. What you see here are the wavelength in different spectral bands from 500 nanometer all the way to 780 nanometers and what you see are the changes as you go through different wavelengths that you can see between the infected leaves and the control leaves and the variation in these spectral features are being captured by the hyperspectral camera. Now these changes are more visible in the late season, we were not able to see these similar changes in the early season as as expected that some of these changes are related to pigmentation in the leaves that is changing and indicating the presence of red blotch in the in the samples. Now here are the spectral features of the both the early stage and the late stage and early stage as I indicated we were not able to see significant changes in the spectral feature all the way from 400 2000 nanometer as shown illustrated here there's no differences between the positive and the negative. Sampled the negative being non infected positive being infected while we were able to see some differences in the spectral around 570 to 6,650 nanometer region as indicated here between the between the positive and the negatives. The positive being infected and the negative being non infected the positive and negative in all our data have been validated using RTPCR as a as a control system to validate the results. Now looking at the hyperspectral based discrimination between impacted and non infected wines in the late stage of the disease, we were able to discriminate based on the spectral features the infected and the non infected. One thing to note is that the separation between them based on the hyperspectral imaging is is much lower than the then the separation we observed using IR spectroscopy. The similar differences could be observed in the hierarchical mean clustering which illustrate that the non infected wines indicated in blue can be discriminated from an infected wines as shown here. But these changes as I indicated were more visible in the late season as compared to the early seasons of the. As measured with IR spectroscopy particularly at the IR spectroscopy. Now, we have also investigated the potential of RGB imaging with data analysis now these are the images that you collect with your cell phone camera with the with the filters so that you can measure the red green and blue channels from the input images. Using deep convolution neural network. The faster our CNN models, we are able to then discriminate between fresh and infected leaf just to illustrate the concept here we have taken this data of infected and non infected leaves train the models and within 20 epochs that 20 cycles of running through the deep convolution neural network we were able to achieve high order of accuracy approaching point nine and also the loss factor decreases significantly as a number of epochs of running the simulation through the to the neural network in this case. So some of the key learnings that we had from this is that mid eye spectroscopy can detect red blotch virus infection in early season. We believe that these are due to differences in the carbohydrate accumulation. In the leaves, which can further be related to the reduction in the car in the sugar in the berries in the in the product in the fruit itself. So the mid eye spectroscopy can detect red blotch viral infection in the early season. We believe that combination of spectroscopy and machine learning models can improve detection accuracy and enable automation, which would be required for field deployment. Near IR hyperspectral and RGB imaging has a potential for late season detection of viral infection potentially we think that these are mostly relative changes in the pigment content. Although evaluation of expanded near our region which ranges all the way from 42500 nanometer may need further evaluation. And similarly deploying cameras that goes from 400 to 2500 nanometer in the hyperspectral region may also enable improve the detection in the early stages. Our detection limit was range from 400 to 1000 nanometer in those regions we were mainly seeing changes in the early in the late season, rather than in the early parts of the season. It's important to keep in mind that various agronomical factors can also influence spectral features so does the training model has to represent the appropriate conditions is a very important. Whatever we train the data to detect, you know, there are other changes that takes place due to, you know, nutrient stress or irrigation status of the wines as we illustrated in our IR spectroscopy data. And we believe that appropriate training model needs to be developed to represent those conditions. With that, we'd like to thank you for your attention. And happy to take questions in the Friday session. Thank you very much.