Field-to-Deployment: Benchmarking AlexNet and DenseNet for Real-Time Banana Leaf Disease Diagnosis

Abstract

Banana (Musa spp.) is one of the most cultivated fruit crops and second largest fruit crop after mango in India. In the field, fungal diseases such as Sigatoka disease (Mycosphaerella musicola) and Cordana disease (Cordana musae) are major limiting factors in the production of bananas and result in large yield and post-harvest losses which have a downstream impact on farmer income and food security. Early diagnosis of such diseases is critical for successful crop management in a sustainable way. Two deep learning architectures, AlexNet and DenseNet are proposed in this study and trained on a custom database of 5163 images of banana leaves captured in the field, which were classified into three classes, Healthy, Sigatoka and Cordana. The test accuracy for AlexNet and DenseNet were 94.9% and 90.8%, respectively, based on accuracy, precision and recall. A mobile application called LeafCheck was created to facilitate real-time disease evaluation at the field level, and deliver information about disease, recommended agrochemical interventions, and precautions validated by local agricultural authorities to the farmer. The proposed system is a practical, scalable, and inexpensive tool that can assist in taking decisions earlier in disease management in banana cultivation, which is a natural resource of great nutritional and economic value.