AI-based Taxonomic Classification ofPollinators and Advanced AoA Estimationfrom mm-Wave and microwave signals

Abstract

Biodiversity is declining at an unprecedented rate worldwide, with insects, particularly pollinators, among the most severely affected groups. Robust monitoring of pollinator populations is essential for safeguarding ecosystem stability, wild plant diversity, and global food production. However, traditional taxonomic identification is labour-intensive, dependent on expert knowledge, and often requires lethal sampling. Moreover, insect abundances fluctuate strongly across space and time due to weather and land-use dynamics, making high-resolution, scalable monitoring challenging.

Although Machine Learning (ML) approaches have been explored for image-based insect identification, their sensitivity to lighting, background clutter, and image quality limits their usability in real-world monitoring. By contrast, Millimetre-Wave (mmWave) radar operating in the 30\,GHz range with wavelength comparable to insect body and wing dimensions offers a non-invasive means to capture fine-scale biomechanical signatures, including micro-Doppler modulations generated by wing flapping. Despite this promise, most radar-based entomology studies focus on morphological parameter estimation (e.g., Wingbeat Frequency (WBF), mass, velocity) and do not exploit the rich harmonic content or modern ML capabilities needed for reliable species-level classification.

This thesis addresses this gap by proposing a non-invasive, low-cost mmWave radar framework for hierarchical taxonomic classification of pollinating insects. As an initial contribution, the thesis investigates the fundamental interaction between flying insects and mmWave antennas through electromagnetic simulation. A dynamic insect model is developed to quantify variations in antenna gain and reflection coefficient ($S_{11}$) induced by micro-Doppler motion in the antenna's near field. These simulations provide initial evidence that insect wing dynamics create measurable perturbations in received mmWave signals, motivating the development of a radar-based classification framework.

Building on this insight, the thesis introduces a mmWave radar pipeline for hierarchical taxonomic classification of pollinating insects. A comprehensive signal-processing and ML pipeline is presented to isolate micro-Doppler activity, extract informative temporal--spectral features, and classify insects at family, genus, and species levels. Explainable Artificial Intelligence (AI) methods are integrated to reveal the critical bio-mechanical and harmonic features underpinning model decisions, improving trust and interpretability.

Beyond empirical radar sensing, the thesis explores the potential of Joint Sensing and Communication (JSC) for insect monitoring. The feasibility of insect detection from a modulated communication signal is also demonstrated, enabled by a semi-supervised segmentation algorithm that autonomously identifies micro-Doppler-active signal regions.

Finally, the thesis contributes to compact antenna design for multi-target localization by addressing different spherical-mode-based Angle-of-Arrival (AoA) estimation methods. A novel Virtual Modes (VM) concept is introduced to enhance angular resolution, reduce computational cost, increase degrees of freedom, and minimise size - weight constraints in compact antenna systems.