Group Members:
Derik Vo
Veda Patel
Yasser Siddiqui
According to Adhikari, Oh, and Panthee (2019) they found that the world wide production of tomatoes is an estimated 161.7 million metric tons. This has an estimated evaluation of %59 billion dollars, of that the USA produces an estimated 13.2 million metrics tons-- placing third in total world production of tomatoes. According to the USDA (2022) late blight has inflicted an estimated $6.7 billion dollars in annual losses that includes tomatoes and potatoes.
The USDA defines blight as a fungus-like pathogen that destroys a plant's leaves, stem, fruit, or tubers. Adhikari et al. (2019) explains blight reproduces asexually, and that it essentially goes dormant in winter and infects plants in warm and humid conditions. They state the condia, or a spore Britannica, usually germinates at 8 to 32 celsius. Once these fungal spores grow they will eventually kill the host cells by producing toxins. These toxins and display symptoms such as "dark, small, necrotic, coalescing, and concentric lesions...on the leaf surface" (Adhikari et al., (2019). Additional, they state that lesions also appear on the stem and can extend to the flesh of the fruit.
Adhikari et al (2019) state the most common method of prevention of blight is through screening. However, they state that this can be labor intensive to identify and treat blight. Because Blight causes such a significant loss in crop loss early detection is paramount. Due to the cost of manual screening we need an effective way to identify the early stages of blight. For our project we will build two neural networks for the purposes of classification. Model one will be classify Tomatoes and will be used to determine if a tomato is healthy or has blight and which stage. Model two will be classify Potatoes and will be used to determine the same information for potatoes.
The model will be evaluated using a confusion matrix, but we will primarily be focusing on the recall score, specifically limiting the number of false negatives. Missing early stages of blight can have disastrous effects on crop yield due to spread. By the time late stage blight has set in, then it will be too late. Therefore our focus will be the recall score of early blight.
For our tomato EDA process and potato EDA process we used Google google Colaboratory. Due to github file size limitations, to reproduce our results users will need to create a short cut from this Google Drive. If that method is not available user's will need to download the tomatoes dataset from the PlanVillage_data Set and zip the classes into a folder with the plants name. After users have created their zip folder they will need to upload it to google drive and mount their drive. The structure of the file path needs to have the master folder as "[plant name]" and sub folders should be "[[class 1]_[class 2]_[class 3]]" for example. Tomato folder should have the sub folders Tomato_healthy, Tomato_Early_blight, Tomato_Late_blight. File names are case sensitive as well.
During the EDA phase we downloaded the tomato dataset from PlantVillage where we found a total of 4500 images the distribution of images is as follow.
| Healthy | Early Blight | Late Blight |
|---|---|---|
| 1000 Images | 1591 Images | 1909 Images |
| Healthy | Early Blight | Late Blight |
|---|---|---|
| 152 Images | 1000 Images | 1000 Images |
After loading our data we wanted to take the average pixel values of each class since the images were separated in their relative classes.
Although not immediately obvious there is already a difference that can be observed in the averaging of the pixel values of the three different classes. The healthy leaves show a bit more of a lighter green and shine. This could be indicative that healthier leaves reflect light better. The early blight leaves averaged out the darkest. This could be indicative that the loss of health lead to less reflectiveness in the leaves. The dark patches were also lowering the average value as well. The Late Blight, however, averaged out even lighter, but the it still doesn't have as much a reflectiveness to it, and the color value started to shift to yellow, presumably from all the green and brown color values from mixing.
The next step was to reorganize the file data set files into two categories one to determine if a plant is healthy or has blight, and two to determine if a plant has early blight or has late blight. These two categories will be used to train the two models mentioned earlier. The following provides insight on the structure of our data.
For our modeling stage we have two separate notebooks one for our Tomato which built 9 models to classify if a tomato is healthy or if its in early or late stages of blight. For our second model for Potatoes we built 5 models.
| Precision | recall | F1 | |
|---|---|---|---|
| Potato early blight | .88 | .77 | .82 |
| Potato late blight | .88 | .94 | .91 |
| Potato healthy | .99 | .99 | .99 |
For our tomato model our best accuracy score was ~96%, for our validation out score was 92%; However, it had a low recall score for early blight (.78). Which is not a result we want because the goal of our model was to limit the number of false negatives for early blight being categorized as healthy.
| Precision | recall | F1 | |
|---|---|---|---|
| Potato early blight | .98 | .98 | .98 |
| Potato late blight | .96 | .97 | .97 |
| Potato healthy | .96 | .80 | .87 |
For our potato model our best accuracy was ~98% for the training set and ~91% for our validation set. This data set had a relatively good recall score (~97%) for early blight; however, it should be noted that there were severe class imbalances. In this set there were only 152 healthy images compared to 1000 early blight images and 1000 late blight images. Because of this imbalance this model may not generalize well when presented with new data. 152 images is also way to low to get any decent training for a solid model. This can be fixed by getting more images in the future, or by creating and saving augmentations and adding them to the dataset.
Our model for tomatoes performs pretty well when measuring for accuracy; however, when measuring for recall the model suffers with incorrectly classifying healthy plants as having blight, but that is not a big issue other than the resources it takes to check up on the plant. Here it also does well in that it does not miss classify a blighted plant as healthy.
When evaluating our model for potatoes it had similar performances to out tomato model. The recall score is slightly higher when classifying potatoes (.96 against .78). However, there was one instance of it miss classifying a late blight potato as being healthy. This is bad because because late stage blight runs has the increase odds of spreading to other plants.
Overall, our models did well when measuring accuracy, but when looking at recall score the score was much lower; however when looking at the confusion matrix we noticed it was mostly miss-labeling diseased plants as healthy which shouldn't be a bit issue
For our next steps we can use our two models to use transfer learning to build a model that incorporates information learned from both plants to help classify if a plant has blight.
Additionally we can build a website that would allow users to submit their images to increase the available data points to train on. This would hopefully address class imbalances as we get more people to crowd source their data.
With this improved model we can pair it with machines to identify when a plant needs intervention. Furthermore, we can also work with engineers to build a machine to automatically apply treatment to infected plants.