Authors: Cooper Moon – Colorado School of Mines
Title: Informing Post-Wildfire Hydrologic Modeling with LSTM-Based Streamflow Models
Presentation Type: Poster
Abstract: Forest disturbances, such as wildfires, increasingly affect large areas of the U.S. and can lead to devastating increases in runoff and sediment transport, posing short- and long-term risks to nearby communities and ecosystems. Predicting post-fire streamflow is also essential for flood mitigation and water resource management yet remains a challenge due to the dynamic nature of disturbed landscapes. For this study, we employ long short-term memory (LSTM) models, a type of machine learning that can identify complex patterns and relationships that traditional models often miss. LSTMs are well-suited for modeling hydrologic response to disturbance given their ability to retain long-term dependencies. We created, and then compared, two LSTMs to simulate post-fire streamflow: one trained on pre-fire streamflow and climate data and another on post-fire data. Preliminary results show that the post-fire-trained model outperforms the pre-fire-trained model for the same post-fire test window for unseen basins, using half the training data (5 vs. 10 years), suggesting that wildfire-induced changes to ecosystems diminish the relevance of pre-fire data for post-fire simulation. Our findings highlight the potential for LSTMs to improve post-wildfire streamflow modeling. This work supports NextGen-compatible model development for the National Water Model by identifying model parameters and data sets that can inform traditional hydrologic models and enhance post-fire streamflow simulations.