johnswyou/wavelet-lstm-camels-models

Wavelet-LSTM CAMELS Streamflow Models

A collection of 61,380 pre-trained LSTM models for daily streamflow forecasting across 620 USGS catchments from the CAMELS dataset.

Each catchment has 99 independently trained models:

• 33 wavelet filters × 3 lead times (1, 3, 5 days) = 99 wavelet-enhanced models
• 33 matching baseline models (same architecture, no wavelet transform)

Models are designed to be ensembled across wavelets for robust predictions with uncertainty estimates.

Dataset Structure

wavelet-lstm-camels-models/
├── README.md
├── filters.json                  # Wavelet filter coefficients (33 wavelets)
├── stations.json                 # Index of all 620 station IDs
├── 01013500/
│   ├── metadata.json             # Scaler parameters + feature selection info
│   ├── leadtime_1/
│   │   ├── db1/
│   │   │   ├── model.keras       # Wavelet-LSTM model
│   │   │   └── baseline_model.keras
│   │   ├── db2/
│   │   │   └── ...
│   │   └── .../                  # 33 wavelet directories
│   ├── leadtime_3/...
│   └── leadtime_5/...
├── 01022500/...
└── .../                          # 620 station directories

File Descriptions

• metadata.json (per station): Contains all scaler parameters (data_min, data_max for feature and target scalers) and feature selection indices for every (leadtime, wavelet) combination. This eliminates the need for sklearn pickle files.
• model.keras: Wavelet-enhanced LSTM model (Input → LSTM(256, tanh) → Dropout(0.4) → Dense(1)). Input features include MODWT wavelet coefficients of hydrometeorological variables.
• baseline_model.keras: Baseline LSTM with the same architecture but using only raw features (no wavelet transform).
• filters.json: Wavelet and scaling filter coefficients for all 33 wavelet families, needed to compute the MODWT transform at inference time.

Model Architecture

All models share the same architecture:

• Input: sliding window of 270 days × N selected features
• LSTM: 256 units, tanh activation, no return sequences
• Dropout: 0.4
• Dense: 1 unit, linear activation (predicts streamflow Q)

Feature count (N) varies per station and wavelet due to automated feature selection (hydroIVS with CMI, tolerance 0.05).

Training Details

• Data: CAMELS daily hydrometeorological data (1980–2014)
• Train/test split: first ~85% train, last ~15% test (temporal split)
• Features: 8 base features (Q, precipitation, temperature, radiation, etc.) expanded to 65 via MODWT (6 wavelet levels + 1 scaling level per feature) plus unix timestamp
• Wavelet transform: Maximal Overlap Discrete Wavelet Transform (MODWT) at 6 decomposition levels
• Scaler: MinMaxScaler (0–1 range) fitted on training data
• Framework: TensorFlow/Keras 2.15

Usage with Python Package

pip install wavelet-streamflow-forecast[hub]

from wavelet_streamflow import Forecaster

# Automatically downloads models for this station from HuggingFace
f = Forecaster("01013500", leadtime=1)
result = f.predict("path/to/01013500_camels.csv")

For selective download (single wavelet):

from huggingface_hub import snapshot_download

path = snapshot_download(
    "johnswyou/wavelet-lstm-camels-models",
    repo_type="dataset",
    allow_patterns=["01013500/metadata.json", "01013500/leadtime_1/db1/*"],
)

Available Wavelets (33)

Wavelet Family	Wavelets
Daubechies	db1, db2, db3, db4, db5, db6, db7
Symlets	sym4, sym5, sym6, sym7
Coiflets	coif1, coif2
Least Asymmetric	la8, la10, la12, la14
Best Localized	bl7
Fejér-Korovkin	fk4, fk6, fk8, fk14
Han linear-phase moments	han2_3, han3_3, han4_5, han5_5
Morris minimum-bandwidth	mb4_2, mb8_2, mb8_3, mb8_4, mb10_3, mb12_3, mb14_3

See the MATLAB Wavelet Toolbox documentation for details on wavelet filter coefficients.

Background

This dataset is the result of research conducted in the following thesis:

You, J. S. (2024). Improving Short-term Streamflow Forecasting with Wavelet Transforms: A Large Sample Evaluation. UWSpace. https://hdl.handle.net/10012/21222

For more on wavelet methods, see:

Percival, D. B., & Walden, A. T. (2000). Wavelet Methods for Time Series Analysis. Cambridge University Press. https://doi.org/10.1017/CBO9780511841040

Citation

If you use these models, please cite the thesis above and the CAMELS dataset:

Newman, A. J., Clark, M. P., Sampson, K., Wood, A., Hay, L. E., Bock, A.,
Viger, R. J., Blodgett, D., Brekke, L., Arnold, J. R., Hopson, T., & Duan, Q.
(2015). Development of a large-sample watershed-scale hydrometeorological data
set for the contiguous USA: data set characteristics and assessment of regional
variability in hydrologic model performance. Hydrology and Earth System Sciences,
19(1), 209–223.

License

MIT — same as the training codebase.

CAMELS Data Attribution

The CAMELS dataset is produced by NCAR and is in the public domain (U.S. Government work). See CAMELS homepage for details.