Monitoring water flows in highly anthropized ecosystems is crucial for assessing eutrophication in water bodies. Ephemeral streams in semi-arid areas present significant challenges for environmental monitoring due to their sporadic flow (2–3 days/year), high intensity, and remote locations (often lacking connectivity and energy sources). Autonomous artificial vision systems deployed at the edge offer a viable solution for this context due to their low cost and high deployment security. However, deploying these systems in different locations requires frequent model retraining to adapt to new environmental conditions, which increases energy consumption and computational resources. This paper proposes an Artificial intelligence of things (AIoT) architecture to identify the presence of water in ephemeral streams, enabling in-place model retraining without compromising the general characteristics of the model. Our results demonstrate that this approach provides a more sustainable and scalable solution for monitoring ephemeral flows, achieving a 64% faster retraining time and a 73% lower energy consumption compared to complete retraining, while maintaining an F1 score around 0.90 for both old and new datasets.