Artificial intelligence with earth observations provides continuous streamflow data across varying wildfire recurrence and recovery scenarios

Abstract

Wildfires continue to alter watershed hydrology globally. Numerous studies have analyzed post-fire hydrology through before- and after-fire comparisons—often focusing on a single fire event. This has led to a significant gap in research, where continuous, daily streamflow data are needed to capture both the short- and long-term effects of recurring fires. Our study, for the first time, demonstrates how Artificial Intelligence (AI) through Machine Learning/Deep Learning (ML/DL) models can fill this gap by harnessing Earth Observations of burn severity, fire radiative power, and burn dates alongside traditional watershed attributes and drought indicators. We present a highly efficient DL model based on the Long Short-term Memory (LSTM) algorithm that is both reproducible and generalizable across multiple fire-affected watersheds in the western United States. Despite watersheds’ diverse climate conditions and markedly different hydrologic characteristics, the model consistently achieves strong performance, with Kling Gupta Efficiency ranging between 0.7 and 0.9. Our primary contribution is testing the model at a daily time-scale to model three distinct fire scenarios: no significant fire history yet the occurrence of a large fire, a longer return interval and no-fire period between phases of recurring fires, and frequent and repeated fire activity. We further investigate whether a relatively simple ML model such as Support Vector Machine (SVM) can achieve comparable accuracy in capturing these complex post-fire scenarios. By offering this fully open-source, generalizable AI solution powered by Earth Observations, newer and better modeling tools can provide post-fire daily streamflow data—a step in the right direction to safeguard water supply reliability and support watershed management across fire-prone regions worldwide.