Gaussian Mixture LSTM Ensemble Framework for Continental-Scale Soil Moisture Prediction over the Contiguous United States

Thie upload contains a Jupyter notebook (GMLSTM_conus_ensemble.ipynb) implementing a Gaussian Mixture Long-Short Term Memory (GMLSTM) network workflow for soil moisture prediction over the contiguous United States (CONUS). It trains and evaluates a probabilistic neural network model that produce mixture-based predictive distributions, enabling both point prediction and uncertainty quantification of soil moisture dynamics. What the notebook does Defines a PyTorch dataset class that loads dynamic time-series inputs, static attributes, and a target variable from user-specified data paths. Computes global mean/std statistics from the training split and standardizes inputs/targets for training (with scalers derived from training data only). Trains a GM-LSTM model using 5-fold cross-validation split by location (KFold with shuffling and a fixed random seed for reproducibility). Saves model checkpoints per epoch within each fold. After training each fold, runs inference on the fold’s held-out locations (using the validation split as the test set in the notebook) and exports per-location time series CSV files containing: date, soil_moisture, lon, lat mixture parameters: pi_1..pi_3, mu_1..mu_3, sigma_1..sigma_3 Input data sources (https://doi.org/10.5281/zenodo.18016491) The modeling framework is designed to ingest harmonized datasets derived from: International Soil Moisture Network (ISMN) in-situ observations ERA5-Land reanalysis meteorological forcings NLDAS-2 land surface forcing fields Daymet daily gridded meteorological data HydroATLAS physiographic and hydrological attribute