Abstract:
Accurate identification of pests and crop diseases is crucial for modern agriculture and the
preservation of the environment. This paper presents G-AgriNet, a state-of-the-art automated
system developed specifically for the classification of agricultural diseases, in response to this
pressing need. The approach we use starts with meticulous preparation of the data, including
normalization, improvement, and a partitioning of 70% for training and 30% for validation sets.
This guarantees a comprehensive evaluation. By using a Graph Attention Network (GAT)
architecture, we may efficiently capture intricate connections within data formatted as a graph.
This is achieved by consistently improving the model via extensive ablation experiments,
resulting in the creation of a dual-layered graph attention network (GAT) with multiple attention
layers. The training utilizes the Adam optimizer, with a batch size of 256 and a learning rate of
0.0001, to achieve optimal efficiency and high accuracy. Utilizing dropout regularization with a
rate of 0.2 mitigates overfitting while maintaining the model's integrity. By using manual features
for image preprocessing, the active learning technique can strengthen the relationship between
nodes and edges in the GNN model, leading to enhanced interpretability and accuracy.
Visualization methods like SubgraphX, GNNExplainer, GraphLIME, Grad-CAM, and Gradient
analysis are used to find important nodes that are necessary for accurately classifying diseases. By
utilizing the CCMT Crop Pest and Disease Detection dataset, our model attains an outstanding
accuracy of 99.25% using a dataset consisting of 96,366 photos. This underscores the significant
influence of modern AI techniques and active learning tactics in agricultural diagnostics. This
research contributes to the area by offering a comprehensive foundation for effective and
environmentally friendly methods for identifying diseases and pests in agriculture.
