Wildfire is an increasingly frequent disturbance that alters factors that mediate catchment hydrologic state, and therefore likely has a significant impact on the timing and magnitude of flow. However, the general effect of wildfire on specific stream flow remains unclear. Part of the difficulty in analyzing the effect of wildfire is the lack of high quality “unburned” controls: hydraulic state is highly dependent on local weather, which varies significantly year to year. In addition, nearby catchments may experience similar weather, but have very different hydraulic state due to variations in vegetation cover, underlying geology, etc. Thus, neither previous years nor nearby catchments serve as ideal controls. However, we wondered if the tools developed by the artificial intelligence (AI) community could help reduce the error present in comparisons between nearby catchments. By comparing the predicted flow from an AI-model to observed flow following a fire, the effect of wildfire on stream flow could be quantified at an accuracy proportional to the accuracy of the model.
Results/Conclusions
We developed an AI model which we trained to predict stream flow. The model consists of an artificial neural network with 50 residual convolutional layers. We trained the model using daily stream flow data available from the USGS Gauges II dataset. We also developed a new method for training the network, which significantly increased its accuracy, which we call the “Fourier L2 Loss.” We then fine-tuned the model on catchments for which wildfire records are available on the Monitoring Trends in Fire Severity database. We were able to achieve a level of accuracy of <5% percent error for over 50 of these streams. We then used the model to predict one year prior, and up to 10 years following wildfires in these streams. We are currently comparing these predictions to the observed flow values following fire to quantify the magnitude and duration of wildfire-induced changes to stream flow.