Lightweight edge AI vision models for IoT-based insect monitoring

Smart automated insect monitoring is essential for early detection of insect pest infestations in orchards. It assists farmers in controlling insect pest populations in their fields and preventing crop losses and improving crop quality. Traditional approaches relying on manual inspections are labor-intensive, error-prone, and highly challenging for large-scale or remote orchards. With recent advances in Artificial Intelligence (AI) and Information and Communications Technology (ICT), there is an opportunity to automate the insect monitoring process effectively and directly in the field. However, key challenges persist, such as limited energy availability, unreliable connectivity in remote areas, data scarcity, and the dynamic nature of orchards. This thesis addresses these challenges through an integrated framework for AI-powered IoT-based resource-constrained edge devices, targeting Brown Marmorated Stink Bug. The work was carried out in three main phases. First, a low-power edge-based detection system was developed and deployed using a Microcontroller Unit (MCU)-based OpenMV board equipped with a camera to explore the feasibility of capturing, processing, and transmitting insect trap images entirely on the edge. This involved optimizing the entire pipeline, including image capture, inference, data transfer, and energy management, while ensuring the system could operate in harsh and remote orchards conditions with limited power and connectivity. The main aim was to validate the concept of performing all image processing directly on low-power consumption, resource constrained hardware and to evaluate its practicality in deployment scenarios. The device operated fully on-device without cloud connectivity and processed two images in ~30 seconds while consuming less than 300 mA of current in active mode. Additionally, this phase enabled the collection of a real-world image dataset under uncontrolled outdoor conditions. In the second phase, the thesis shifts toward efficient Deep Learning (DL) model design, building on the real-world data obtained in the first stage. A lightweight DL model, named SemiY-Net, was proposed for simultaneous object segmentation and counting. The model was designed to meet the strict memory and computational constraints of MCUs while providing meaningful outputs. Several strategies were employed to achieve this balance, including architectural pruning, layer-level memory profiling, and skip connection management. Unlike typical segmentation models designed for high-resource platforms, SemiY-Net introduces a design that deliberately reduces spatial resolution specifically at the output, trading off accuracy in favor of deployment feasibility. The model was trained and validated on a dataset of annotated insect trap images, supporting both segmentation masks and insect count labels. In addition to its accuracy, it runs entirely on-board MCUs without reliance on a server or cloud support. Quantitatively, SemiY-Net reduced parameters by >75% compared to state-of-the-art models, required <1 MB of storage and RAM, and achieved ~5 J energy consumption per inference on an OpenMV. On the field dataset collected in Phase 1, it achieved Dice Similarity ~85% and counting Mean Squared Error ~1.32. The final stage of the thesis addresses a fundamental challenge in deploying DL models in real-world scenarios, particularly the difficulty of collecting and labeling proper datasets in remote and dynamic environments. To overcome this, a self-improving and fully edge-based framework was developed where the MCU-based edge nodes are supported by an edge server (e.g., Raspberry Pi) that handles DL model retraining and model adaptation. The MCUs perform real-time inference using the proposed SemiY-Net and transmit only compressed Regions of Interest (RoIs) to the edge server, reducing communication overhead by ~84%. The edge server then reconstructs the raw captured images from the received RoIs, generates the corresponding pseudo- labels and masks, and fine-tunes SemiY-Net to adapt to new field conditions and data distributions. In addition to the reconstructions, artificial images were also synthesized from the RoIs to increase the training sample diversity. These reconstructed and augmented samples were then used in periodic continual learning cycles, where the server fine-tuned SemiY-Net with the new data and redistributed updated model weights back to the edge nodes. This system enables continuous adaptation without human intervention. Quantitatively, models fine-tuned with pseudo-labels retained accuracy comparable to those trained with expert annotations, while inference energy consumption on MCUs remained ~5 J and communication energy consumption dropped >80% compared to raw image transfer. Together, these three phases form a practical edge AI-based IoT system for insect monitoring. The system covers the entire pipeline, from hardware to model development and deployment. By eliminating reliance on cloud and human annotation, it delivers intelligent automation directly to the orchard, enabling reliable and long-term insect monitoring even in resource-constrained environments.

Grantor dc:publisher

University College Cork

Year dc:date.issued

2025

Author dc:creator

• Kargar, Amin

Advisors dc:contributor.advisor

• O'Flynn, Brendan
• Tedesco, Salvatore

• TinyML
• Edge AI
• Incremental learning
• Computer vision
• Embedded systems
• Image segmentation
• Deep learning
• Insect monitoring
• Smart agriculture

Statement dc:rights

• © 2025, Amin Kargar Barzi.

Licence dc:rights.uri

• https://creativecommons.org/licenses/by/4.0/

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10468/18840

OAI identifier oai:identifier

oai:cora.ucc.ie:10468/18840