Closing the Yield Gap: A Scalable Computer Vision Pipeline for In-Field Detection of Devastating Banana Diseases

Although bananas are planted due to food security of the world, they are subject to numerous fungi, bacteria and viruses that reduce the production. Visual method is the normative way through which banana plant infections are identified, and this technique lacks scalability, and it is arbitrary. A new method of deep learning that enables automatic, fast, and accurate identification and diagnosis of the pathogens that affect banana leaves is presented in this paper. Our hybrid Convolutional Neural Network (CNN) model was developed and compared with EfficientNetB0 and ResNet50. The model has been trained using a diverse set of images and augmented even more using data augmentation methods. The model achieved great results of 98.7 percent precision and F1score of 0.986. It was capable of locating Black Sigatoka, Fusarium Wilt and Banana Xanthomonas Wilt. Gradientweighted Class Activation Mapping (Grad-CAM) has validated that the model identifies biologically significant lesion areas, which enhance interpretability. The findings indicate that the framework is an effective and well-defined precision agriculture tool that simplifies easy response to quick action and possibly effective in identifying diseases in various crops.