Mobile-Friendly Plant Disease Detection and Classification using Metadata and Asymmetric Knowledge Distillation

The increasing interest in the mobile applications of real-time plant disease detection systems has given rise to new difficulties in building computational models that can perform with high accuracy with limited resources. This paper presents an inventive concept that leverages metadata and asymmetric knowledge distillation to improve the use of mobile applications for detecting and classifying plant diseases. It is reported that a teacher model (ResNet – 50) trained on environmental factors (e.g. temperature, humidity, and soil) enriched dataset was converted to a student model (MobileNet) which is mobile deployable. The knowledge embedded in soft labels of the teacher model was distilled and taught to the student model allowing it to practice high accuracy data classification with less computations. Adding metadata further gives evidence to the assertion and helps improve the overall detection accuracy. The results of the experiments indicate the effective use of knowledge distillation in a mobile-based plant disease classification where a lightweight MobileNet model performs comparably with the ResNet-50 model. This structure offers a practical and scalable solution for real-time disease monitoring and identification, allowing farmers to make timely decisions with a smartphone.