Irrigation Detection

A deep learning pipeline for detecting outdoor irrigation usage from smart water meter (AMI) data. Given hourly consumption readings, the model separates total water usage into irrigation and non-irrigation components.

How It Works

The pipeline uses a DilatedUNet1D — a 1D U-Net with dilated convolutions in the bottleneck for long-range temporal context. The model takes 14-day (336-hour) windows of hourly consumption and directly predicts the irrigation component at each hour.

Input: Hourly water meter readings (meter ID, timestamp, reading value) Output: Per-hour decomposition into total, irrigation, and non-irrigation readings

Quick Start

From source

pip install .
irrigation-detect \
    --input data/readings.csv \
    --model models/best_irrigation_model_336h.pth \
    --output results.csv

With Docker

docker build -t irrigation-detection .
docker run --rm \
    -v $(pwd)/data:/app/data:ro \
    -v $(pwd)/models:/app/models:ro \
    -v $(pwd)/output:/app/output \
    irrigation-detection \
    --input /app/data/readings.csv \
    --model /app/models/best_irrigation_model_336h.pth \
    --output /app/output/results.csv

With Docker Compose

# Place your input CSV in ./data/ and model in ./models/
docker compose run --rm detect

Input Format

A CSV (or Parquet) file with three columns:

Column	Type	Description
meter_id	string	Unique identifier for the meter
timestamp	datetime	Hourly reading timestamp (UTC)
reading	float	Consumption value for that hour

Example:

meter_id,timestamp,reading
M001,2024-01-01 00:00:00,0.12
M001,2024-01-01 01:00:00,0.08
M001,2024-01-01 02:00:00,0.05

Readings should be hourly interval values (consumption during that hour), not cumulative register readings. Values should be non-negative.

Output Format

Column	Type	Description
meter_id	string	Meter identifier
timestamp	string	ISO-formatted timestamp
total_reading	float	Original consumption value
irrigation_reading	float	Predicted irrigation component
non_irrigation_reading	float	Predicted non-irrigation component

Model

The DilatedUNet1D architecture:

• Encoder: 4 blocks of dilated convolutions with max pooling (1 → 16 → 32 → 64 → 128 channels)
• Bottleneck: Dilated convolutions (dilation 2, 4) for multi-scale temporal context (256 channels)
• Decoder: 4 upsampling blocks with skip connections
• Output: ReLU activation (irrigation is non-negative)
• Window: 336 hours (14 days)

The model was trained on synthetic mixed-use meter data. You must supply your own trained model weights (.pth file).

Configuration

All settings are configurable via CLI flags or environment variables:

CLI Flag	Env Var	Default	Description
--input	INPUT_PATH	—	Path to input CSV/Parquet
--model	MODEL_PATH	—	Path to model weights (.pth)
--output	OUTPUT_PATH	—	Path for output CSV
--device	DEVICE	cpu	Torch device (cpu or cuda)
--batch-size	GPU_BATCH	2000	Windows per inference batch
--source	SOURCE_LABEL	—	Optional label for source column

Development

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run on example data
irrigation-detect --input examples/sample_data.csv --model path/to/model.pth --output out.csv

Citation

If you use this tool in your research, please cite:

California Data Collaborative (2026). Irrigation Detection Pipeline.
https://github.com/California-Data-Collaborative/irrigation-detection

License

MIT License. See LICENSE for details.