Predicting changes in river channel conveyance and geometry using a machine learning approach

Abstract:

Changes in river channel conveyance (i.e. the discharge carrying capacity) directly impact flood inundation frequency, navigability and the overall stability of riverine infrastructure. However, there are no operational methods for robustly predicting changes in conveyance over time. Historically, many studies have assessed changes in river channel geometry by using historical time series of stream measurements made at river gauges. However these transect measurements are affected by numerous uncertainties such as the measurement location, which is not always well documented in historical archives. There is thus a need to develop an automated method that can filter data archives and unlock predictions of changes in river channel conveyance over time.

This paper reports on the development of the first artificial intelligence (AI) enabled automated, open-source workflow for evaluating changes in river channel conveyance using streamflow measurements. To ensure consistency of measurement location (and facilitate detection of changes in channel capacity), we first implement a series of unsupervised machine learning based clustering methods that detect the most consistent measurement location over time at any stream gauge. We assess a variety of machine learning algorithms for clustering measurement location, including a Density-Based Spatial Clustering of Applications with Noise (DBSCAN); a Gaussian Mixture Model (GMM); and spatial clustering (kMeans). We assess the skill of our location-detection method using both objective evaluation metrics and visual assessment of stream channel data across the coterminous United States. Then, we develop a supervised machine learning based data processing pipeline to predict changes in channel geometry and conveyance at-a-gauge. We use lagged streamflow, precipitation and temperature as predictors, and evaluate the predictability over different lead times and seasons.