Code and data for "Transformers Improve Streamflow Prediction in Karst Watersheds by Capturing Long-Term Memory Effect of Storage"

This repository contains code and data for Choi et al. (under review). Author: Seohye ChoiLast modified: 12/18/2025 # Code Transformer-RR.ipynb: Jupyter notebook for finetuning, training, validation, and interpretative analysis of a simplified Transformer model, using datasets described below. LSTM-RR.ipynb: Jupyter notebook for finetuning, training, and validation of a baseline LSTM model, using datasets described below. # Data karst.mat: Inputs and target variables for training and validation, compiled for nine snow-dominated karst watersheds in the Western U.S. The first dimension represents each watershed in the following order: 1. Little Bighorn River2. Salina Creek3. Steptoe Creek4. Lamoille Creek5. Mission Cr. ab Reservoir6. Cache Creek7. Logan River8. Red Butte Creek9. Blacksmith Fork non_karst.mat: Inputs and target variables for training and validation, compiled for nine snow-dominated non-karst watersheds in the Western U.S. The first dimension represents each watershed in the following order: 1. Halfmoon Creek2. Black Gore Creek3. Andrews Creek4. Minam River5. Rock Creek6. Bobtail Creek7. Lake Fork8. Stillwater Fork9. Headwater Weber River For both datasets, the third dimension represents 1. Streamflow (m/d),2. Liquid water input (i.e., snowmelt plus rain (m)),3. PET (mm). # Reference Choi, S., Tennant, H., Hill, D., Neilson, B., Newell, D., Ashmead, N., McNamara, J., & Xu, T. Transformers Improve Streamflow Prediction in Karst Watersheds by Capturing Long-Term Memory Effect of Storage. Under Review.