Distributed Deep Learning in IoT Sensor Network for the Diagnosis of Plant Diseases

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Early plant disease detection is crucial for agricultural productivity and food security.
• Centralized deep learning models face challenges in IoT-based agricultural environments due to data transmission and computational demands.
• Existing methods are often unsuitable for large-scale, decentralized agricultural deployments.

Purpose of the Study:

• To propose a distributed deep learning framework using Federated Learning (FL) for plant disease diagnosis in IoT sensor networks.
• To evaluate the effectiveness of a hierarchical FL approach compared to a standard pipeline for multicrop disease classification.
• To address the limitations of centralized models in resource-constrained agricultural settings.

Main Methods:

• Developed a distributed deep learning framework integrating IoT nodes and an edge computing node.
• Implemented the Federated Averaging (FedAvg) algorithm for collaborative model training without local data transfer.
• Evaluated two training pipelines: a standard single-model approach and a hierarchical model combining crop classification with crop-specific disease models.
• Utilized an EfficientNet B0 model for image-based plant disease diagnosis.

Main Results:

• The hierarchical FL approach demonstrated improved per-crop classification accuracy and robustness against environmental variations.
• The standard FL pipeline offered lower latency and reduced energy consumption.
• Experimental results were validated on a multicrop leaf image dataset with realistic augmentation.

Conclusions:

• Federated Learning provides a viable solution for decentralized plant disease diagnosis in IoT agricultural settings.
• Hierarchical FL architectures enhance classification performance and resilience in diverse agricultural conditions.
• The choice between standard and hierarchical FL pipelines depends on balancing accuracy, latency, and energy efficiency requirements.