CropNet: A Smart Framing Approach to Soybean Disease

Background: Rapid spread of plant diseases is still one of the serious threats exerted against global crop production, which causes significant economic losses and puts food security at risk. Several studies have investigated deep learning architectures (CNN/R-CNN variants, EfficientNet and hybrid models) that were developed using big datasets viz PlantVillage, hyperspectral images and field-collected samples. Promising results have indicated that these approaches would achieve classification accuracies up to 96.52%. Nonetheless, problems including poor diversity of dataset, lack of generalization to real-world situations, computations involved and deployment limitations do exist. Methods: This article summarizes several recent research works, which examine various deep learning models to detect plant diseases. It also presents evaluations of CNN-based models, region-based detection techniques (R-CNNs), as well as lightweight high-performance architectures that include EfficientNet and hybrid deep learning method variants. The considered models were trained on popular datasets: PlantVillage, hyperspectral image databases and data collected from the field. Moreover, this paper introduces our 16 layers CropNet model trained and tested on a self-collected dataset and then compared to the proposed deep learning architectures. The comparative methodological study considers accuracy, computational efficiency, generalizability and deployment feasibility, leading to a comprehensive evaluation of the proposal in cropnet architecture. Result: The accuracy of the proposed CropNet model outperforms other networks and achieves performance close to the state-of-the-art with a value of 98.75%. Yet, the findings also point to current limitations due to dataset variety, generalization on realistic field scenarios and computation costs.