ABSTRACT
Wheat serves as one of the most important staple cereal crops worldwide, and as such, its health has a direct bearing on food security. However, diseases such as Septoria tritici blotch and Stripe Rust cause considerable yield losses in wheat crops. Traditional methods for detecting diseases in wheat often rely on visual inspections, which are usually time-consuming, labor-intensive, and subjective, with considerable chances of error. Deep learning has ushered us into an era of enhanced automated diagnostics for wheat diseases; the methodology is becoming ever more plausible and dependable. To improve feature representation, the pipeline begins by adjusting brightness and contrast and employing rotations that are dynamically altered throughout training. The suggested preprocessing pipeline was investigated against the backdrop of a Custom Skin Colour (CSC) Model based on ResNet-50, which was improved for multiclass classification with three classes: Rust, Septoria, and Healthy. Accuracy, precision, recall, and F1 score were among the performance metrics that this adaptive preprocessing significantly outperformed when compared to static, fundamental preprocessing procedures. When using adaptive preprocessing approaches to identify disease in wheat, the overall accuracy improved by around 5.7%, from 86.4% to 92.1%. Additionally, mildly afflicted crops showed greater precision and F1 score. In the case of agricultural activities, it is crucial to have a preprocessing pipeline that is always changing. A scalable pipeline that can identify diseases in other food crops, not only wheat, is suggested by the existing constitutive model's flexibility, which will support productive agriculture overall.
