The motivation of this project is totally based on the target user of gardening and the severity of the leaf disease. It is quite clear that gardening has become a common practice for urban life. Nowadays, especially in cities we can hardly find a roof that has no plant in it. But still, we cannot neglect the fact that most of the urban people involved in gardening has no prior knowledge about farming. They do not know what are the diseases the plants can face, how to detect these diseases and finally how to remedy them. Hence, they often fail to detect the disease in time. Also, leaf disease can be both infectious and non-infectious depending on the nature of a causative agent. If not detected early it can lead to the death of the plant which could have easily been solved by a touch of technology. But most of the features of agricultural technology focuses mainly on field agricultural crops rather that residential gardening. The reason why we are taking this project is to help these urban people get acquainted with a technological tool that will help them to detect the plant diseases in real time.
In our project, our model is a CNN model. A workflow diagram displaying the complete process of the method is shown below:
Figure 1: Block diagram of workflow
In the diagram, we can see that our first step is collecting a suitable dataset which we can use for training and testing our model. The images from the dataset have to be pre-processed before we can use it for training and testing. After training the model, we can test the model by making predictions on some images and validating the prediction with the actual image and also find the accuracy.
Collecting dataset is our first step and also one of the most important step. Before collecting a dataset, there are some features that has to be explored. Since our project requires an image dataset, we have to look at features like the number of image samples, variations of image, background, image format and resolution.
We have collected our dataset from here which contains .jpg type of image. There are ten classes of five different types of plants containing 2273 pictures of both healthy and diseased leaves. All the images in the dataset has similar background. The resolution of all images are also same.
Before using our images for training, the images have to be pre-processed. Pre-processing is required to increase the quality of input image so that the performance of the model can be improved. We followed the following steps for pre-processing our images:
Figure 2: Block diagram of image pre-processing
From the above diagram, we can see that after obtaining the image from the dataset, we will be resizing the images. Then we will normalize the images and finally apply image augmentation.
We started our pre-processing with image rescaling and resizing. Our images from the dataset were 6000x6000 resolution. We resized the image into 256x256 size to ensure that all images stay at a consistent scale.
After resizing the images, we normalized the images. Normalization keeps all the input pixels on similar distribution. In our model, we performed normalization by dividing the pixel values by 255. By dividing the pixel values, we rescale the values between 0 and 1.
For our model, we performed normalization implicitly in the resizing and rescaling step.
Our final step of pre-processing stage is image augmentation. Data augmentation is applied to the existing dataset to artificially increase the size of the dataset by performing various transformations like flipping, distorting, rotating, zooming etc. Data augmentation helps to create variation and lowers the chances of overfitting.
In our model, we have randomly flipped the images horizontally and vertically. We have also rotated the images randomly up to 0.2 radians.
Our Machine Learning model is a CNN model. Our model can be divided into two parts, Feature extraction and image classification. We have used CNN for both feature extraction and image classification.
Feature extraction transforms raw data into numeric values compatible
with machine learning
algorithms. These numeric values can be processed by keeping
the integrity of the information that is found from the original
dataset. For our project, we used CNN for feature extraction. There are
two main layers of feature extraction in a CNN model. If we take a look
at the CNN architecture,
Figure 3: CNN model architecture of feature extraction
Here we can see that there are two layers in feature extraction, Convolutional layer and Pooling layer.
Convolutional layer helps to detect various features and pattern which are present in the image by applying convolutional filters to the input image. This filters then learn extracting different visual features like textures.
Pooling layer selects the minimum value within a specific region of the image and down samples it. This helps to retain the relevant information of the image by reducing the relative dimension.
By stacking these two layers, the model can learn how to identify features from the input image.
Our model also has these two layers. We have used two convolution layers and two pooling layers. The convolution layers of our model have 32 filters and each filter has 3x3 kernel size. We have used the activation filter ReLU (Rectified Linear Unit) which is a popular activation function. ReLU uses the formula,
f(x) = max(0,x)
Which means that the output can be found only if the input value is positive, otherwise it will be zero. If there is any negative input, ReLU deactivates the neuron by converting it to zero[]. As a result, exponential growth in computation can be prevented. Our two pooling layers capture the most important features with a pool size of 2x2.
In a CNN model, the classification layer determines the predicted value on the activation map. Once the model has extracted features, it needs to classify those features into learn the features of the data.
CNN classification also has two layers. If we look at the classification architecture,
Figure 4: CNN model architecture of classification
We can see that there is a flatten layer and a dense layer.
The flatten layer flattens the 2D layers from feature extraction into 1D vector. This layer converts the features into a format which can be processed by the dense layers.
After flattening, all the layers are passed to the dense layer or fully connected layer. Dense layer contains multiple neurons which are fully connected to each other. This layers learns to classify the extracted features into their corresponding classes.
For image classification, we have also used these two layers. The
flatten layer flattens the 2D feature maps into a 1D vector before going
to the dense layer. Our first dense layer is a fully connected layer
consisting of 64 units or neurons. This layer has ReLU activation
function which applies linear
transformation to the input data. Our second dense layer is the output
layer of the model. We have used softmax activation function here. Each
output value of this layer represents the probability of input of a
particular class.
Figure 5: Model summary
After preprocessing and building the model, our model is ready for training. We have split our dataset into two partitions. The first partition is our training dataset. Which is 80% of the total dataset. And the remaining 20% is the testing dataset. We further split the testing dataset into two more partitions. Which are test and validation. Each partition is 10% of the testing dataset.
We started training our model with the train dataset. The validation dataset is used to evaluate the model's performance. We used the loss function Sparse Categorical Crossentropy. This function calculates the cross-entropy loss between the true labels and the predicted labels. The value increase when the difference increases. We also used Adam optimizing algorithm. Adam can adapt the learning rate of each parameter also maintaining a moving average of squared gradients. As a result, Adam provides stable and fast optimization.
At first, we started training our model with a batch size of 32. We also had an epoch value of 50, which is the number of iterations the model will go through over the entire training dataset. With this values, our did not have a stable accuracy. So, we had to tune the parameters. We tried different batch sizes, epoch values, and learning rate of the optimizer and found the best accuracy with a batch size of 32, epoch 30 and learning rate of 0.001.
After completing the training part, we tested our model and made predictions. At first, we evaluated the model on the test dataset and found the loss and accuracy values. For prediction, at first we selected one image sample from test dataset and used the model to make predictions on it. The model was able to make correct predictions. Then we selected 9 images from the test dataset and made predictions. This time, all the predictions were mostly correct. We also displayed the confidence level by taking the maximum predict probability.
Confusion matrix is one of the most common performance measurement used in machine learning. It shows the summary of a model's performance on a test data. This matrix uses the true labels and the predicted labels to show four types of information. These are True Positive (TP), True Negative (TN), False Positive (FP) and False Negative (FN)
Figure 6: Confusion matrix
From the confusion matrix, we can also calculate some important data for each class which we can use to analyze the model. Which are Accuracy, Specificity, Sensitivity, Precision and F1 score.
Accuracy shows the outputs that are actually correct. It is calculated using the formula,
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃 + 𝑇𝑁
Specificity is the model's performance on identifying true negative cases, which is calculated using the
following formula,
𝑇𝑁
𝑇𝑁 + 𝐹𝑃
Sensitivity is the model's performance on identifying true positive cases. Sensitivity is also known as
recall. This is calculated using the formula,
𝑇𝑃
𝑇𝑃 + 𝐹𝑁
Precision shows how many true positives are predicted correctly. Which means finds how many true
positives are actually true positives. The formula which is used to calculate precision is,
𝑇𝑃
𝑇𝑃 + 𝐹𝑃
F1 score can be calculated from precision and Recall. It is the harmonic mean of the precision and recall
[11], which is calculated by,
2𝑇𝑃
2𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁
We have calculated all these values for each class of our model from the Confusion Matrix, and got the following scores,
| Class Name | Accuracy | Specificity | Sensitivity | Precision | F1 Score |
| Guava Diseased | 0.973 | 0.981 | 0.846 | 0.733 | 0.785 |
| Guava Healthy | 0.822 | 1.00 | 0.870 | 1.00 | 0.931 |
| Jamun Diseased | 0.822 | 0.983 | 0.973 | 0.925 | 0.948 |
| Jamun Healthy | 0.733 | 0.990 | 0.809 | 0.894 | 0.850 |
| Lemon Diseased | 0.995 | 0.995 | 1.00 | 0.800 | 0.888 |
| Lemon Healthy | 0.991 | 1.00 | 0.909 | 1.00 | 0.952 |
| Mango Diseased | 0.991 | 0.989 | 1.00 | 0.941 | 0.969 |
| Mango Healthy | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
| Pomegranate Diseased | 0.982 | 0.980 | 1.00 | 0.851 | 0.920 |
| Pomegranate Healthy | 0.982 | 1.00 | 0.862 | 1.00 | 0.925 |
Table 1: Model evaluation matrices
We also have two graphs which can show the performance of our model. The first Graph is the Training and Validation accuracy graph. This graph shows the progression of accuracy of the model over each epoch. The training accuracy is the accuracy of the model on the training dataset and the validation accuracy is the accuracy on the validation dataset which is different from the training dataset.
Figure 7: Training and validation accuracy graph
From the graph, we can see that the accuracy starts to increase as the model learns to adjust the parameters. We can also see that the training and validation curves are increasing at similar level which shows that the model is learning well.
Our other graph is Training and Validation loss graph. This graph shows the change loss of the model over each epoch. The goal here is to minimize the amount of loss over each epoch.
.png "Training and validation loss graph")
Figure 8: Training and validation loss graph
In this graph, we can see that the loss amount starts to decrease with each epoch. The training and validation curves are also on similar position.