Watershed Symposium 2019

Modeling Fire-Induced Changes to River Status Flow With Deep Learning

Summary:
The potential effect of wildfire on Utah’s rivers is poorly understood. Because fires are becoming more frequent and severe, understanding their effect on river “status flow” will be an important component of Utah’s future water security. Using deep learning, we developed a powerful model for predicting flow/fire relationships.

Full Abstract:
Watershed disturbances such as wildfire can fundamentally alter water flow and water chemistry, affecting downstream ecosystems and societies. Predicting the response of streamflow to wildfire is difficult but increasingly urgent in the western U.S. given the observed increases in wildfire frequency, uncertainty in precipitation, and growing human water demand. Much of the difficulty in stream-flow prediction arises not from a lack of data (flow and geological data are available for thousands of catchments), but from a lack of capacity to extract information from these data that is socioecologically relevant. Mechanistic modelling of streamflow-wildfire response based on catchment characteristics has only been partially successful because of substantial variation in catchment characteristics including geological features deep below the surface, fire-induced biological changes on the surface, and meteorological variability. Recent advances in machine learning, particularly in Artificial Neural Networks (ANNs) may have opened the door for capturing this complexity. ANNs are a special type of machine learning algorithm inspired by human learning. In an ANN, groups of artificial neurons interact and learn together, eventually allowing the network to perform tasks that would otherwise be impossible for a computer. The structure of neurons in an ANN allows it to abstract information from complex signals and explore interactions between sub-features, potentially allowing prediction of ecologically complex phenomena such as wildfire-flow response. Here, we used an ANN to analyze USGS flow data from over 4,000 watersheds in the contiguous US, 600 of which experienced a wildfire in the past 30 years. This ANN was able to predict alterations to the “status flow” (i.e. flow characteristics of flow regime) of streams affected by fire, including modification of the amount and timing of flow as well as timeframes of recovery following wildfire. We explore how this machine learning approach could usher in a new status quo in extracting maximum information from flow stations and predicting response of western watersheds to wildfire. Data-driven models that more accurately represent likely scenarios can help ensure water security for Utah’s communities in the face of rapidly accelerating changes in climate, fire frequency, land use, and resource scarcity.