Hybrid Multi-Objective Neural Architecture Search for Lightweight Patch-Based Mistletoe Classification in UAV Imagery

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Automated remote sensing for vegetation monitoring is hindered by complex structures and cluttered backgrounds.
• Current vision frameworks for detecting parasitic plant infestations use overparameterized Convolutional Neural Networks (CNNs), limiting edge computing deployment.
• There is a need for efficient and accurate CNN architectures for hardware-constrained monitoring devices.

Purpose of the Study:

• To propose a novel method for automatically designing lightweight CNN architectures.
• To address the efficiency-accuracy trade-off in vegetation monitoring and parasitic infestation detection.
• To enable real-time digital image processing on edge devices.

Main Methods:

• A two-phase hybrid multi-objective Neural Architecture Search (NAS) strategy was implemented.
• The Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) minimized classification error and trainable parameters.
• An Iterated Local Search (ILS) metaheuristic refined solutions using aerial RGB imagery and a balanced dataset.

Main Results:

• A 10-layer CNN topology with high feature-extraction efficiency was discovered.
• The optimized model achieved high performance metrics: Accuracy (0.979), F1-Score (0.979), Precision (0.982), Recall (0.976), and Jaccard Index (0.958).
• The lightweight CNN operates with only 2040 trainable parameters, outperforming existing models.

Conclusions:

• The developed NAS strategy successfully created a lightweight and efficient CNN architecture.
• The optimized CNN is highly viable for real-time digital image processing on hardware-constrained monitoring devices.
• This approach advances automated remote sensing for vegetation monitoring and pest detection.