Online Agenda
•
• 1 min read

Building LSTM and Transformer Models for Hydrologic Prediction

This tutorial provides a hands-on introduction to applying deep learning models—LSTM and Transformer—to hydrologic prediction using the CAMELS dataset. Participants will learn how to preprocess data, construct and train models, and evaluate their performance on rainfall–runoff tasks. No prior hydrology experience is required—just curiosity about deep learning for Earth science.
Jiangtao Liu
Written by Jiangtao Liu

📘 Overview

In this tutorial, we will explore how to use the CAMELS (Catchment Attributes and Meteorology for Large-sample Studies) dataset to build deep learning models for rainfall–runoff prediction. You will learn how to preprocess hydrologic data, construct and train LSTM and Transformer models, and evaluate their predictive performance.

To make the tutorial lightweight and runnable on Colab (even without GPU), we’ll use a small subset (20 basins) of CAMELS that has been pre-processed and stored as a NetCDF file.

💡 The full CAMELS dataset can be downloaded and processed using the provided Python scripts (01.download_camels.py, 02.prepare_camels.py), which takes about 20 minutes.

Resources

References

Liu, J., Bian, Y., Lawson, K., & Shen, C. (2024). Probing the limit of hydrologic predictability with the Transformer network. Journal of Hydrology, 637, 131389.

Liu, J., Shen, C., O’Donncha, F., Song, Y., et al. (2025). From RNNs to Transformers: Benchmarking deep learning architectures for hydrologic prediction. EGUsphere, 2025.

Jiangtao Liu
Written by

Jiangtao Liu

I am interested in using multiple satellite datasets, in-situ observation datasets, and reanalysis products to investigate how climate variability and human activities affect water resources. My approach integrates physics-based hydrological models with deep learning techniques, ensuring that model predictions remain both accurate and physically interpretable. In parallel, I develop BERT/GPT-based foundation models that can be fine-tuned for tasks such as streamflow forecasting, soil moisture prediction, and water quality assessment. Ultimately, I aim to deliver robust, scalable, and transparent modeling frameworks that guide decision-making from local watersheds to global scales, helping mitigate risks such as droughts, floods, and landslides in a changing climate.

Start the conversation

Related

See all Online Agenda
Online Agenda
•

Talk to Your ML Models. Insights from Sensitivity Analysis (XVARS) and Explainable AI (XAI)

Sensitivity Analysis (XVARS) and Explainable AI (XAI)

Banamali Panigrahi
Written by Banamali Panigrahi
Online Agenda
•

Large Language Model (LLM) Agent in Hydrology

This tutorial introduces AQUAH, the first language-based agent specifically designed for hydrologic modeling. With simple natural-language prompts (e.g., 'simulate floods for a specific basin'), AQUAH autonomously retrieves data, configures hydrologic models, executes simulations, and produces analyst-ready reports, simplifying complex environmental modeling tasks.

Zhi Li
Written by Zhi Li
Online Agenda
•

Web Application Development for Hydrologic Modeling and Visualization

This session introduces Tethys Platform, an open-source framework for building earth-science- focused web applications. For over a decade, scientists have used Tethys to build web applications for hydrologic modeling, visualization, and decision support. Participants will see two complementary approaches to web app development.

Nathan Swain, Shawn Crawley, and Corey Krewson
Written by Nathan Swain, Shawn Crawley, and Corey Krewson