BNNDC: Branched neural network for plant disease identification

Deep learning (DL) advancements have contributed to the success of vision-based tasks for solving real-world problems. DL applications in agriculture are increasing as researchers find it valuable for developing solutions to ensure global food security. However, commonly used DL architectures were designed for large datasets comprising millions of hand-annotated images. Such large, annotated datasets are limited for agricultural applications. Further, specific agricultural applications, such as disease identification, present a non-trivial problem due to similar visual features of different diseases and background noise. Hence, developing lightweight and taskspecific DL architectures may help improve plant disease identification performance. This study proposes a novel neural network architecture named Branched Neural Network for Disease Classification (BNNDC)-Net. The performance of the proposed BNNDC-Net was compared with large (DenseNet169 and ResNet50) and small (ResNet18 and MobileNetV2) conventional DL architectures by training and testing on publicly available and field-acquired data. BNNDC helped eliminate false positives compared to conventional DL models. BNNDC-Net and BNNDC-Res with 3.86M and 6.81M parameters, outperformed ResNet50 for disease identification with 85.60% and 73.36% fewer trainable parameters and 13.93% and 14.26% improvement in testing accuracy, respectively. All trained DL architectures were additionally tested for generalization capability on field images. BNNDC-Res outperformed all other architectures by at least 5.17% in generalization accuracy. This study demonstrates the potential of a lightweight but limited data-requiring DL architecture for plant disease identification and generalization to field conditions. Future work will further expand the study to diagnose diseases for other field crops.