Integrating Quantized Edge AI and LoRaWAN for Visual-Based Wildlife Detection in Connectivity-Limited Regions
Year: 2026
Type: conference-paper
Abstract: This work investigates an end-to-end Edge-AI deployment pipeline tailored for connectivity-limited environments, addressing both model-centric and data-centric design considerations. On the modelling side, we adopt the nano variant of the YOLOv26 architecture to fine-tune a region-specific endangeredspecies detector, and evaluate multiple quantization strategies to characterize the trade-offs between predictive performance, latency, and memory usage. On the data-transfer side, we emulate a LoRaWAN to assess its suitability for daily image transmission, examining the behavior of its Adaptive Data Rate mechanism under varying compression levels and spreading factors. Experimental results show that dynamic quantization provides an effective balance of efficiency and accuracy, delivering a latency of 27 ms per image classification and a compact model size of 1.57 Mb, whilst still achieving F1-scores of up to 97%. Furthermore, our LoRaWAN simulations indicate that, with image compression ratios between 30% and 90% and transmission distances up to 750 m, reliable daily data transfer is feasible.
Institution: Monash University, Indonesia
Language: English
Topic: IoT and Edge/Fog Computing
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204234048