AUTHOR=Prince Reazul Hasan , Mamun Abdul Al , Peyal Hasibul Islam , Miraz Shafiun , Nahiduzzaman Md. , Khandakar Amith , Ayari Mohamed Arselene 

TITLE=CSXAI: a lightweight 2D CNN-SVM model for detection and classification of various crop diseases with explainable AI visualization

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 15 - 2024

YEAR=2024

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2024.1412988

DOI=10.3389/fpls.2024.1412988

ISSN=1664-462X

ABSTRACT=Plant diseases have an adverse effect on crop productivity and quality, making them a danger to world agriculture. However, identifying and categorizing these diseases can be a time-consuming and error-prone task. This is where our suggested technique comes into play. By using deep learning-based algorithms and convolutional neural networks (CNNs) in the identification of leaf diseases in four crops that are important to the economy: strawberries, peaches, cherries, and soybeans, these four important crop leaf diseases can be accurately identified and categorized.The purpose of this work is to categorize 10 classes of diseases for peach, cherry, soybean, and strawberry crops, with 6 diseased classes and 4 healthy classes, using a deep learning-based CNN-SVM model. For this purpose, some pre-trained models were also trained. The accuracy range is 53.82% to 98.8% for VGG16, VGG19, DenseNet, Inception, MobileNetV2, MobileNet, Xception and ShuffleNet. On the other hand, the proposed model's average accuracy is 99.09%.The proposed model and VGG16 pre-trained model have similar accuracy, but the important thing about the proposed model is that its trainable parameter is so much lower than the others, which makes the model better and unique. That is why the CNN-SVM model has been chosen over VGG16 and other models. Additionally, the proposed model performs very well, as evidenced by its 99% F1-score, 99.98% value of Area Under the Curve (AUC), and 99% value of precision.Class activation maps were also created using the Gradient Weighted Class Activation Mapping (Grad-CAM) technique to visually explain the disease that the proposed model detected and a heatmap was created to show the region that needs to be classified.