1. Exploratory Data Analysis (EDA)

First, we begin by importing essential libraries and modules required for deep learning in image analysis tasks.

from tensorflow import keras
import matplotlib.pyplot as plt 
from keras import backend as k
import numpy as np
import os
import cv2

Subsequently, moving to proceed with conducting exploratory data analysis.
We encode the names of the plant species using label encoding technique. This process involves converting the names of the twelve classes into corresponding numerical representations.

folder_train = ‘path_to_ train’
class_folders_train = sorted(os.listdir(folder_train))

Output:
class_folders_train = ['Alstonia Scholaris diseased (P2a)', 'Alstonia Scholaris healthy (P2b)', 'Arjun diseased (P1a)', 'Arjun healthy (P1b)',  'Bael diseased (P4b)',  'Basil healthy (P8)', 'Chinar diseased (P11b)', 'Chinar healthy (P11a)', 'Gauva diseased (P3b)',  'Gauva healthy (P3a)', 'Jamun diseased (P5b)', 'Jamun healthy (P5a)', 'Jatropha diseased (P6b)', 'Jatropha healthy (P6a)', 'Lemon diseased (P10b)', 'Lemon healthy (P10a)', 'Mango diseased (P0b)', 'Mango healthy (P0a)', 'Pomegranate diseased (P9b)', 'Pomegranate healthy (P9a)',  'Pongamia Pinnata diseased (P7b)', 'Pongamia Pinnata healthy (P7a)']

label_mapping = {}
for i, class_folder in enumerate(class_folders_train):
label_mapping[i] = class_folder
display(label_mapping)

Let's explore the result of this encoding process by examining the transformed data.
{0: 'Alstonia Scholaris diseased (P2a)',
1: 'Alstonia Scholaris healthy (P2b)',
2: 'Arjun diseased (P1a)',
3: 'Arjun healthy (P1b)',
4: 'Bael diseased (P4b)',
5: 'Basil healthy (P8)',
6: 'Chinar diseased (P11b)',
7: 'Chinar healthy (P11a)',
8: 'Gauva diseased (P3b)',
9: 'Gauva healthy (P3a)',
10: 'Jamun diseased (P5b)',
11: 'Jamun healthy (P5a)',
12: 'Jatropha diseased (P6b)',
13: 'Jatropha healthy (P6a)',
14: 'Lemon diseased (P10b)',
15: 'Lemon healthy (P10a)',
16: 'Mango diseased (P0b)',
17: 'Mango healthy (P0a)',
18: 'Pomegranate diseased (P9b)',
19: 'Pomegranate healthy (P9a)',
20: 'Pongamia Pinnata diseased (P7b)',
21: 'Pongamia Pinnata healthy (P7a)'}

In the next step, we conduct data exploration by randomly choosing one image from each species and plotting it to provide a summarized visual representation. This step allows us to gain an initial understanding of the characteristics exhibited by each plant species.

Figure 1: Random image samples from twelve plant species

To plot this, you can use the following lines of code:

visualized_data = []
for i in range(len(class_folders_train)):
    image_name = random.choice(os.listdir(folder_train + class_folders_train[i]))
    image = cv2.imread(os.path.join(folder_train, class_folders_train[i], image_name))
    visualized_data.append(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))

fig = plt.figure(figsize=(20, 15))
columns = 5
rows = 5
for i in range(1, columns*rows):
    if (i < len(visualized_data)+1):
        image = visualized_data[i-1]
        fig.add_subplot(rows, columns, i)
        plt.title(class_folders_train[i-1])
        plt.imshow(image)
plt.show()

In addition, we created a bar chart to represent the number of plants in two distinct states: "diseases" and "healthy”. To do this, we need to count the number in two states of each species:

class_folders_train_temp = [x for x in class_folders_train if x not in ['Bael diseased (P4b)', 'Basil healthy (P8)']]
diseased  = []
healthy  = []
for i in range(len(class_folders_train_temp)):
    class_i = class_folders_train_temp[i]
    num_class_i = len(os.listdir(folder_train + class_i))
    if (i % 2 == 0):
        diseased.append(num_class_i)
    else:
        healthy.append(num_class_i)

plant_species.append('Basil')
healthy.append(len(os.listdir(folder_train + 'Basil healthy (P8)')))
diseased.append(0)

plant_species.append('Bael')
healthy.append(0)
diseased.append(len(os.listdir(folder_train + 'Bael diseased (P4b)')))

Figure 2: Health Condition Distribution: Plants in Diseased and Healthy States

This chart aims to provide a comparison by illustrating the distribution of twelve plant species across two states of health conditions. The bars representing the "diseases" state indicate the number of plants within each species that are affected by diseases or experiencing some form of health issues. The bars representing the "healthy" state indicate the number of plants within each species that are in good health and free from any noticeable diseases or health problems. At first glance, we can observe that the distribution of the two states within the same species is imbalanced. Furthermore, the number of samples across different plant species is also uneven. Additionally, Basil and Bael are two specific cases worth noting: Basil does not have any samples in the diseased state, while Bael does not have any samples in the healthy state.

2. Network Details

The neural network created in this case uses the Tensorflow architecture. The VGG-19 [2] neural network is used.

Figure 3: VGG-19 architecture

After completing the model training process, we will utilize the trained model to generate predictions on both the validation set and the test set. This step allows us to thoroughly assess the performance and accuracy of the model. The accuracy of validation set and testing set provide valuable insights into the model's ability to generalize to unseen data and its overall effectiveness in solving the given problem.

# Load the model that is saved after the training process
model.load_weights("path_to_model_checkpoint")

# Evaluate the model on the valid set
valid_loss, valid_accuracy = model.evaluate(valid_generator, steps=len(valid_generator))
print("Valid Loss:", valid_loss)
print("Valid Accuracy:", valid_accuracy)

# Evaluate the model on the test set
test_loss, test_accuracy = model.evaluate(test_generator, steps=len(test_generator))
print("Test Loss:", test_loss)
print("Test Accuracy:", test_accuracy)

Output:
4/4 [==============================] - 17s 4s/step - loss: 0.6710 - accuracy: 0.7636
Valid Loss: 0.671012818813324
Valid Accuracy: 0.7636363506317139
4/4 [==============================] - 17s 4s/step - loss: 0.6784 - accuracy: 0.7455
Test Loss: 0.6784456968307495
Test Accuracy: 0.7454545497894287

Now, let's examine the results. The results indicate that the model achieved an accuracy of 76.4% on the validation set and 74.5% on the testing set. Additionally, the F1-score obtained on the test set is 0.7376. Considering the complexity of multiclass classification involving 22 classes and potential data imbalance, this F1-score is considered to be reasonably good. It indicates that the model effectively achieves a balance between accurately identifying positive samples (precision) and appropriately capturing positive samples (recall).

Figure 4: Confusion Matrix for 22 Classes on Testing Set