Optimizing Multi-Level Crop Disease Identification using Advanced Neural Architecture Search in Deep Transfer Learning

Abstract

Efficiently managing crop diseases holds immense potential for optimizing farming systems. A crucial aspect of this process is accurately identifying infection levels to enable targeted and effective disease treatment. Despite recent advancements, developing a reliable system for identifying and localizing crop diseases in complex, unstructured field environments remains challenging. Such a system requires extensive annotated data. This study comprehensively evaluates deep transfer learning techniques for identifying the degree of rust disease infection in Morocco's Vicia faba L. production systems. A vast dataset captured under natural lighting conditions and various crop growth stages was created to facilitate this research. Ten deep learning models were rigorously assessed through transfer learning, establishing a benchmark for this task. Deep transfer learning achieved high classification accuracy, with F1 scores consistently surpassing 90.0%. Training time for all models was reasonably short, under 2.5 hours. The NVIDIA Quadro P1000, known for its exceptional performance, was pivotal in achieving this outcome. The Neural Architecture Search-based model emerged as the top performer, achieving an impressive overall F1 score of 90.84%. Three models achieved F1 scores near or above 90.0%, highlighting the effectiveness of deep transfer learning for rust infection identification. This research illuminates the potential of deep transfer learning in detecting and diagnosing crop diseases, specifically rust infection in Vicia faba L. production systems. The findings contribute to developing robust disease management strategies, improving agricultural practices, and enhancing crop yield.