Deep-Neural-Networks-for-Crop-Disease-Detection-and-Interpretability-XAI

Dataset Analysis:

The dataset used in this thesis comprises a collection of images representing different plant species and their associated diseases. The dataset is divided into four main categories: corn, potato, rice, and wheat. Each category consists of specific disease classes and a healthy class for comparison. The total number of classes in the dataset is 14, with a combined total of 13,024 images.

A. Plant Overview

1. Corn Plant Species: The corn category contains images depicting diseases such as Common Rust, Gray Leaf Spots, and Northern Leaf Blight, along with a class representing healthy corn plants. The dataset includes a total of 3,852 images, with Common Rust has 1,192 images, Gray Leaf Spot having 513 images, Northern Leaf Blight has 985 images, and the Healthy class has 1,162 images. These images were obtained from the widely-used PlantVillage dataset.
   

2. Potato Plant Species: The potato category comprises images related to Early Blight, Late Blight, and healthy potato plants. A total of 2,152 images are included in this category, with 1,000 images each for Early Blight and Late Blight, and 152 images representing the Healthy class. The images in this category were collected from the PlantVillage dataset.
   

3. Rice Plant Species: The rice category includes images representing diseases like Brown Spot, Leaf Blast, and Neck Blast, along with a Healthy class. The dataset consists of 4,078 images, with 613 images for Brown Spot, 977 images for Leaf Blast, 1,000 for Neck Blast, and 1,488 for the Healthy class. The images in this category were sourced from the ”Dhan�Shomadhan” dataset [4] and the ”Rice Leafs” dataset from Kaggle.
   

4. Wheat Plant Species: The wheat category contains
   images related to diseases such as Brown Rust, Yellow Rust, and healthy wheat plants. It consists of 2,942 images, with 902 images for Brown Rust, 924 images for Yellow Rust, and 1,116 images representing the Healthy class. The origin of the wheat images used in this dataset was Kaggle.
   B. Data Distribution
   I can calculate the distribution of images in each class to understand the relative representation of different plant diseases and the healthy class.
   Total images: 2942 images

My methodology:

This section presents the proposed methodology of the proposed methodology for exploring deep neural networks for crop disease detection and interpretability. Our model has two parts. The first is dedicated to capturing visual attributes, and the second is focused on providing information about our model’s predicted effects and potential biases. We will now provide a breakdown of each individual step within this architectural framework,

A. Input Image
In our Dataset, there is no directory for the training, testing, and validation. All the image classes stack into one folder. So before sending the images to our model, we need to split the dataset into three parts

I) Data Splitting: Data is often separated into three sets: training, testing, and validation. The training set is used to train the model, the validation set is used to fine-tune hyperparameters and model selection, and the test set is used to evaluate the model’s performance on unknown data. So, we split the dataset with 70% for training, 20% for testing, and 10% for validation.

II) Data into Batch: Crop disease images from the training dataset are presented to the proposed model batch by batch. For our work batch size is 32.

B. Image Preprocessing
Preprocessing is a data preparation step in machine learning that aims to improve data quality and suitability for training models. It involves techniques like removing unwanted distortion, balancing class imbalances, and applying various transformations to images. It involves various techniques to clean, transform, and balance the data. So, we ensure that all images are the same size and to enable more effective processing, we pre-processed each image by scaling it to 2992993. The rescaling technique is used to reduce the image pixels will be scaled between 0 and 1. Our model can learn more interesting features why we also use different types of augmentation techniques such as

I) Color Jitter: Color jitter contains different types of image color modification tools like brightness, contrast, saturation, and hue. That means we send the data to our model with variety .

II) Random Flip: Images may be flipped horizontally or vertically at random. This is done at random to show the model diverse perspectives on items. It’s analogous to showing the computer how things may seem from the opposite side. This improves the computer’s learning since it sees more ways that objects might appear.

III) Random Height and Weight: Images may be set height and weight at random. This is done at random to show the model diverse perspectives on items. This improves the computer’s learning since it teaches models to detect even if height and weight are changed randomly of image.

III) Random Rotation: Using random rotations causes variety in the orientation of the pictures. This is especially beneficial when images in the collection have varying orientations.

IV) Random zoom: Using random zoom causes variety in The visualization of the pictures

C. Feature Extraction
In our projecThe we were tasked with interpreting and making sense of the information hidden inside images. To address this, we used a complex mechanism known as transfer learning, which allows us to apply the insights obtained from training neural networks on large and varied picture datasets.
We began by selecting three well-known pre-trained convolutional neural network (CNN) models: InceptionV3, VGG19, and Exception. These models have been painstakingly trained on enormous arrays of photos to detect nuanced patterns, characteristics, and relationships. Because of their thorough training, they have already learned to recognize distinct objects, forms, textures, and colors - key features critical for image comprehension.
Instead of building our models from scratch, we used the power of these pre-trained models. We input our images into these neural networks, which then analyzed the images layer by layer, unraveling their contents. The models divided the photos into smaller, relevant components and highlighted the distinctive traits that distinguish each image.
So, it can be said that we obtained a significant advantage by incorporating these pre-trained models into our strategy. We were able to concentrate on the finer aspects of our images that were important to our research, such as specific features that may assist us in distinguishing between distinct classes and concepts. Finally, we ran all the models with same hyperparameter setup like
batch size = 32, iterations= 30000, learning rate = 1e-4 and epoch= 105, the epoch is calculated by this formula iterations/Len(train dataset)/batch size
We also use the optimizer as Adam and the loss function as ’categorical cross entropy for multiclass classification. We also, use early stop criteria at the time of training to avoid overfitting and we save the best model for further activity.

D. Evaluate The Model:
Accuracy, Precision (P), recall (R), and weighted F1-score are used to compare performance. The model’s misclassification rate has been utilized as one of the metrics to effectively compare its performance across several classes. To evaluate how well the model performs, we utilize the weighted F1- score measure.

E. Apply XAI for Explanation:

We went a step further in our research by using eXplainable Artificial Intelligence (XAI) approaches to the models. When we examined the images from our dataset, this allowed us to understand how these models were operating. Grad CAM [24], which means Gradient-weighted Class Activation Mapping, was used as our first method. With the use of this method, we may identify the regions of the image that the model is focusing on more than others.
But, while we were looking at the data, we saw something interesting. For most of the images we tested, our model seemed to zoom in on the right parts – the important areas of the pictures. This matching between where the program was looking and the correct parts of the pictures showed that our program was doing a good job at understanding and finding the important things.