Given the previous epochs from 1 to 8, has values for their F1 Score and Accuracy. It shows that there is possible room to improve the model with more fine-tuning and possibly an early stop in training to not have the model over-train to get the result of 1.0 for F1 Score and Accuracy. Another possible reason for the result could be the large imbalance of images between classification labels and that the Plant Village data set is considered to be a small data set. These types of results are good but not realistic for model prediction power. Even though these values are incredibly high, the training and validation loss is still fairly low and slowly converging to 0. This means that the model is learning over time, which is a good sign and shows room for improvement. A possible future improvement for this project would be to find another data set with more specific plant disease information to train the model on. Efforts will be made to look for data that includes more plant pests. That variation would be beneficial because the data set for this project mainly contains crop disease images. The disease Spider Mite, is labeled as a disease in this data set but in reality, is not a disease. It is the only classification label that has pest-inflected crop images in this data set. Having more data on crop pests would be beneficial because crop pests also cause a significant amount of crop loss and damage as well. Technical improvements for this project include developing a stronger data augmentation technique. Instead of using the set transform method which artificially augments the data set,garden grow bags using another package that can actually create the separate images and add them to the data set instead would be interesting to see how it would perform. More Hyper-parameter fine-tuning could be done such as exploring the learning rate and adding an optimizer. More Epochs should be tested in order to see how the model will perform over a greater period of time.

Fewer epochs will also be tested to see how early-stopping the model from training will perform. Another idea is to explore more the Training-Validation-Testing splits chosen for the data. Exploring the performance between different splits could show how to better improve the classification model. To achieve our expected agricultural need of feeding 10 billion people by 2050, we must prioritize minimizing crop loss wherever possible. More research is needed to help develop more tools to assist crop growers with preventing crop loss. The study “Mobile phone use is associated with higher smallholder agricultural productivity in Tanzania, East Africa” by Amy Quandt et al. looks into the relationship crop growers have with their cell phones as agricultural tools to help increase crop yields. “A key result is the positive association between phone use for agricultural activities and self-reported agricultural yields”. Cell phones are increasing accessibility to technological tools that help with agriculture. These technologies for assisting with crop loss will not only be utilized by commercial crop growers or the average hobbyist and enthusiast as well. Whether crop growers use a ViT image classification model or a convolutional neural network image classification model, or another type of machine learning architecture is used, more research is needed to help develop tools to assist crop growers worldwide in eradicating crop loss everywhere! The original Specialty Coffee Asscn. of America Coffee Taster’s Flavor Wheel was developed in 1995 by Ted Lingle, before many advances and methods in sensory science had been developed . To revise this longstanding industry tool, sensory science and statistical methods were applied as novel flavor wheel construction techniques. Even today, across food and beverage industries, very few flavor wheels exist that were created using a scientific approach and a sensory lexicon. A lexicon is a list of vocabulary developed using sensory descriptive analysis methods used to describe a product, along with descriptions of each attribute and reference preparation instructions .

Some notable existing flavor wheels have been created using sensory lexicons, for products such as beer, wine, tea, spices, and even drinking water, but the flavor wheel construction methods differed from those used in the current study Suffet and others 1999; Gawel and others 2000; Koch and others 2012; Lawless and others 2012. Lawless and others used similar statistical techniques to those used in this study to develop the McCormick Spice Flavor Wheel; however, the data used were simply a subset of descriptive analysis data gathered from lexicon development, with no sorting task. In the development of a tea flavor wheel, Koch and others used all descriptive analysis data to perform principal component analysis to determine the positioning in the flavor wheel, but no clustering techniques or sorting exercises were utilized. Gawel and others did use a sorting task, as in this study, for mouthfeel attributes in wine, but slightly different clustering statistical techniques were used, and only 9 experts participated in the sorting exercise. Prior to this project, SCAA and World Coffee Research worked with sensory scientists, industry representatives, and trained panels of judges at the Sensory Analysis Center at Kansas State Univ. and Texas A&M Univ. to develop a lexicon of about 110 attributes to describe flavor , texture/mouthfeel, and amplitude . The WCR Sensory Lexicon was then sent to UC Davis to be sorted into categories and levels to be converted into an updated Coffee Taster’s Flavor Wheel. The words in a flavor wheel serve to standardize training and aid in education and discussion. The original Coffee Taster’s Flavor Wheel has served as a communication tool about coffee products among all components of the industry, including tasters, plants, retailers, exporters and importers, producers, baristas, and consumers. The new Coffee Taster’s Flavor Wheel will serve as an improved communication tool, as it is an organized visualization of the WCR Sensory Lexicon . This tool is the 1st step toward enabling the coffee industry to identify and characterize specific flavor changes and relate these changes to specific variables in the coffee process, which brings us one step closer to understanding which factors drive coffee flavor.Although this type of scientific conversion from lexicon to flavor wheel was unprecedented, existing sensory and statistical methods were also applied for the purposes of this study.

A rapid sensory profiling method sans tasting, called single free sorting, was utilized to determine the similarities and dissimilarities among the 99 coffee flavor attributes. Once the data from the individual sorting tasks were collected and summarized, 2 multivariate statistical techniques were applied. First, to determine the major categories, subcategories, and levels, agglomerative hierarchical clustering was used. Conjointly, to determine the arrangement of these categories and subcategories in the wheel structure, multidimensional scaling was used. AHC and MDS are both techniques used in sensory science to observe and visualize the similarities between different products, consumers, or attributes . In this way, existing sensory and statistical methods were adapted to create a novel method for constructing a flavor wheel from a defined lexicon.Twenty-nine trained descriptive analysis panelists were contacted and recruited from other descriptive studies already in progress at UC Davis. These panelists were not required to be trained specifically on coffee, but they were required to be regular coffee drinkers, had all participated in descriptive analysis on products with complex flavors, and had worked with and been exposed to most of the flavor attributes in the WCR Sensory Lexicon. Panelists were not further trained for this experiment,tomato grow bags because group discussions may have allowed the more opinionated panelists to influence the decisions of the quieter panelists. For this experiment, it was decided to simply allow panelists to draw on their individual experiences and subsequently compile and average all the data, rather than holding group discussions and coming to a group consensus. Recruitment, screening, and scheduling were done via email. Once accepted into the study, participants were sent written instructions to perform the sorting task on the web app remotely from their personal computers. The entire process was online and remote. In order to accurately reflect the coffee industry needs, create an additional set of data, and add more statistical power, 43 coffee industry experts recruited by SCAA performed the same online procedure as the UC Davis panelists. The industry panelists came from all areas of the coffee industry and they all had experience as coffee tasters, but not all of them had experience in sensory descriptive tests.Before the sorting task began, the WCR Sensory Lexicon was reduced to 99 flavor attributes, removing all attributes not exclusively referring to flavors. Specifically, the attribute “astringent” and all attributes in the Texture/Mouthfeel and Amplitude sections were removed. The word sort procedure was originally done in a Steinberg study , to be used as a tool for semantic analysis, particularly regarding connotations. This sorting method has since been adapted to food samples for sensory analysis .

Traditionally, panelists are asked to sort food products or other samples into as many clusters or groups as they choose, in a way that makes sense to them . In free multiple sorting, a rapid sensory descriptive method, panelists repeat this procedure until they feel they have exhausted the sorting possibilities, and then they are asked to provide descriptions for each group of samples . In a study comparing single sorting to multiple sorting, Rosenberg and Kim found that multiple sorting was superior in representing all possible dimensions of categorization of the data. Additionally, one drawback to using single sorting is that the individual data need to be summed together in order to analyze it, so individual data are lost . In this experiment, instead of sorting food samples themselves, panelists were asked to sort the attributes into categories and subcategories without tasting samples and therefore based on their experience and expectations of these flavor descriptors. Thus, this sorting exercise was similar to the original word sort procedure performed by Steinberg in 1967. Sorting the words themselves was appropriate in this case, due to the ultimate goal of using the Coffee Taster’s Flavor Wheel as a tool for coffee industry professionals. Since there was no tasting, fatigue, adaptation, and carryover effects did not bias the data . Additionally, as there were 99 attributes, to avoid fatigue, instead of repeating the procedure multiple times, the panelists each only sorted the lexicon once. A user-friendly web interface was created using AngularJS to allow for simple, efficient sorting of the 99 flavor attributes. This helped to minimize the clutter of index cards and catalyze the data collection process, because data could be stored immediately in Firebase , a free database and web application hosting service. The website had a welcome page and the participants would log in and be greeted with brief instructions and a “begin” button. The users would then see further instructions and the list of attributes, each with Table 3–RV-coefficients from multiple factor analysis comparing UC Davis and coffee industry participants. Industry UCD MFA Industry 1 0.414 0.832 UCD 0.414 1 0.850 MFA 0.832 0.850 1 a question mark to the far right with a scroll-over pop-up with the WCR definition of that attribute. If a user was unclear about the meaning of one of the words of the lexicon, he/she could scroll over the information bubble to access the definition. The participant was able to click and drag the attributes into categories and subcategories, for as many hierarchical levels as they deemed necessary. Once the user felt the attributes were adequately sorted into categories and subcategories, they would press “submit” and the results were immediately stored in the Firebase database.The methods of AHC and MDS were used to represent attribute–attribute relations instead of product–attribute relations, because there were no coffee samples, only attributes. To organize the raw data, a program was written in Ruby to translate the sorting data into matrices that could be used for analysis. For both of these methods, 1st, 2 binary matrices were created for each participant , one for “sibling–sibling” relationships, in which the attributes appeared in the same subcategory, and one for “parent–child” relationships, in which one attribute appeared in a subcategory under another attribute.