Sistema AIoT basado en Edge Computing para el Monitoreo de Recursos Hídricos

Abstract

Abstract: This research presents the development of a flood early warning system, integrating Edge Computing, Artificial Intelligence, and IoT-based alerts. Given the lack of infrastructure and monitoring technology in Ecuador, a decentralized architecture is proposed, employing the self-supervised foundational model DINOv3-ConvNeXt Tiny as a backbone for semantic feature generation. The segmentation head is trained on this model, achieving a mIoU of 0.929 on independent test data. During deployment, the system processes images locally on a Raspberry Pi 4 device, eliminating cloud dependency for model inference. The delimitation of alert points is virtually defined through community knowledge. Alerts are efficiently transmitted via Telegram, optimizing bandwidth usage, while an average RAM consumption of 64.67% ensures system stability. The results demonstrate that, despite an average inference latency of 12.5s and a high average CPU load (99.2%), the solution is technically viable and serves as a foundation for future developments.