ABSTRACT:

This workshop introduces machine learning for hydrologic applications and works through a reproducible modeling workflow in PyTorch. The first half covers ML fundamentals: the supervised, unsupervised, and reinforcement learning paradigms, common task types and algorithm families, and the practical limitations that affect real projects, such as data quality, overfitting, and interpretability. The hands-on portion builds an LSTM model to predict streamflow from the CAMELS dataset, using 10 gauged basins in Southern Appalachia. Participants load and explore the data, then preprocess it with attention to leakage: a temporal train/validation/test split, normalization computed on training data only, log-transformed streamflow, and 365-day input sequences. The model is trained with Adam and a masked MSE loss, then evaluated on a held-out 2008–2014 test period using per-basin NSE scores, hydrographs, and residual plots. The second half turns to model improvement. After diagnosing whether the architecture or the underlying data is the limiting factor, the workshop compares four variants on the same test basins: a baseline LSTM, a version augmented with static basin attributes, a deeper two-layer LSTM, and a causal-convolution ConvLSTM. Throughout, the emphasis is on reading the full distribution of per-basin performance rather than a single average, and on deciding what to change next based on evidence.

Acknowledgements:
This research was supported by the Cooperative Institute for Research to Operations in Hydrology (CIROH) with funding under award NA22NWS4320003 from the NOAA Cooperative Institute Program. The statements, findings, conclusions, and recommendations are those of the author(s) and do not necessarily reflect the opinions of NOAA.