Using Mobile Edge AI to Detect and Map Diseases in Citrus Orchards.

Area of Science:

• Agriculture 4.0
• Computer Vision
• Edge AI

Background:

• Deep Learning (DL) models show promise in Agriculture 4.0 for tasks like disease detection but are computationally intensive.
• High processing demands and poor connectivity hinder DL deployment on mobile devices in field settings.
• Edge computing offers a solution by enabling on-device processing for real-time applications.

Purpose of the Study:

• To propose and validate an edge AI application for detecting and mapping citrus orchard diseases.
• To identify low-footprint DL models suitable for mobile edge devices for fruit detection and disease classification.
• To evaluate the performance of various AI algorithms for both detection and classification tasks in a real-world agricultural context.

Main Methods:

• Compared YOLO and Faster R-CNN for fruit detection in trees.
• Evaluated lean AI models including MobileNetV2, EfficientNetV2-B0, and NASNet-Mobile for image classification.
• Utilized statistical parametric models and a genetic algorithm to analyze disease spread from detection results.

Main Results:

• YOLO demonstrated faster performance than Faster R-CNN for fruit detection with similar accuracy.
• MobileNetV2 and EfficientNetV2-B0 achieved 100% accuracy in disease classification, outperforming NASNet-Mobile (98%).
• MobileNetV2 showed superior timing performance over EfficientNetV2-B0 and NASNet-Mobile for classification.

Conclusions:

• The proposed edge AI pipeline effectively enables mobile, on-device disease detection and mapping in citrus orchards.
• Optimized AI models like MobileNetV2 are crucial for developing efficient edge AI solutions in Agriculture 4.0.
• The system facilitates the assessment of disease spread, supporting agricultural management decisions.