StrawberryWatch: A Temporal Graph Neural Network-Driven Anomaly Monitoring System for Urban Watersheds

StrawberryWatch is a real-time anomaly detection system based on Temporal Graph Neural Networks (Temporal GNN), designed to monitor the ecological health of the Strawberry Creek urban watershed. The system integrates meteorological data and sensor telemetry data to enable intelligent pollution early warning, addressing the issues of long cycle and slow response in traditional manual sampling monitoring, and providing technical support for urban watershed management.

Urban watersheds serve functions such as drainage, landscape, and ecological regulation, but in the process of urbanization, pollutants from stormwater runoff and accidental leaks threaten stream ecosystems. Traditional monitoring relies on manual sampling and laboratory analysis, which has a long cycle and slow response, making real-time early warning difficult. Strawberry Creek is located near the University of California, Berkeley, and is an important local urban water system. The StrawberryWatch project aims to build an automated and intelligent ecological monitoring system for this scenario.

The project uses Temporal Graph Neural Networks (Temporal GNN) for anomaly detection, with significant advantages:

1. Graph Structure Modeling: Model Strawberry Creek as a directed flow graph, with 5 sensor nodes connected according to the direction of water flow, accurately reflecting the pollutant propagation path;
2. Temporal Feature Learning: Capture temporal dependencies such as daily and seasonal cycles of water quality parameters (temperature, conductivity, etc.) through LSTM layers;
3. Anomaly Detection Mechanism: Trigger alarms based on reconstruction errors, without the need for a large number of labeled anomaly samples. The production model DuskCrayfish architecture: The spatial layer (GCN) handles the spatial relationships between nodes, the time layer (LSTM) models dynamic evolution, and the adaptive threshold is adjusted according to rainfall to reduce false alarms.

The system's data comes from two main channels:

• Sensor Telemetry: Sensors along the line collect high-frequency data (minute-level) such as temperature, conductivity, pH value, turbidity, and dissolved oxygen in real time;
• Meteorological Data: Integrate data from the Lawrence Berkeley National Laboratory weather station (air temperature, dew point, humidity, etc.) to include external driving factors. Data fusion embodies best practices in environmental modeling, considering both the monitored object itself and external influencing factors.

The project supports multiple operation modes:

• Training Mode: Pull data from the past 30 days to build a flow graph, train the model and save weights (used for cold start or major updates);
• Update Mode: Fine-tune with new data based on existing weights (daily maintenance);
• Inference Mode: Evaluate the latest 48 hours of data to detect anomalies (triggered manually or scheduled);
• Real-Time Monitoring: Run 24/7, perform inference every 15 minutes, and send email alerts when anomalies are detected.

Experience in deploying similar systems:

• Data Access: Supports public REST API (recommended) and MySQL database connection;
• Model Cold Start: Need to train first to establish a baseline; it is recommended to accumulate 2-4 weeks of data before starting real-time monitoring;
• Alarm Fatigue Management: Adaptive threshold + rainfall detection to reduce false alarms, and visual reports provide context;
• Scalability: The GNN architecture naturally supports adding sensor nodes; updating the graph structure is sufficient without redesigning the model.

StrawberryWatch is a typical application of AI for Earth, with significance including: timely detection of pollution events to shorten response time, reduction of labor costs to achieve round-the-clock monitoring, accumulation of long-term data to support ecological research and decision-making, and provision of technical demonstration for urban watershed management. The project is open-source, providing a reusable framework for other communities, and promoting the transformation of urban water environment management from 'post-event governance' to 'pre-event prevention'.