Towards efficient stream monitoring: A systematic approach for model selection and continuous improvement in Tiny Machine Learning applications

Measuring ephemeral stream flows is essential for ecological and hydrological studies. However, their intermittent nature and remote locations pose challenges for conventional monitoring methods, which often consume excessive energy to capture rare events. We address this with BODOQUE (Bimodal Observational Device for Optimizing Quantification of Ephemeral streams), a dual-mode system that leverages Tiny Machine Learning (TinyML) on low-power microcontrollers. The system remains in an energy-saving sensing state and activates high-precision measurements only when water flow is detected. We present a model selection methodology that balances detection accuracy with inference cost, enabling reliable operation within hardware constraints. To enhance adaptability in diverse environments, we developed a specialized component that facilitates dataset expansion through new field samples. This supports ongoing retraining to maintain model performance under changing conditions. A comprehensive evaluation using real-world data demonstrates that our system can achieve up to 97% annual energy savings compared to traditional continuous monitoring approaches.