Classifying Low Streamflow Parameters in Coastal Southeast Alaska Using Game Cameras and Random Forest Models

June 16, 2023

Kari Lanphier
Master’s Thesis (16 June 2023). DOI:


Southeast Alaska is located on the traditional territory of the Lingít, Haida, and Tsimshian People. It is comprised of the largest temperate rainforest in the world, with subregions receiving over 500 cm of rain annually. Climate change is predicted to alter the region's timing, type, and magnitude of precipitation and increase temperatures year-round. Potential impacts from these changes include decreased snow accumulation at lower elevations and less precipitation in the summer. These changes could result in an increased frequency of low streamflow occurrence across Southeast Alaska. Low flow conditions could impact fisheries, hydropower generation, and municipal water supply in Southeast Alaska. With much of the region being remote, there are limited streamflow gaging stations, which hampers efforts to monitor changing streamflow conditions. To address this paucity of data, this thesis considers two methods to better understand streamflow response to changing climate conditions in Southeast Alaska. Chapter 2 tests the efficacy of using game cameras to collect streamflow parameters during the summer months, and Chapter 3 leverages currently available data to understand the primary drivers of low flows. Chapter 2 successfully collected streamflow parameters of magnitude, timing, frequency, and duration of low and high streamflow events during the summer of 2022. Stage height data (collected from pressure transducers) and pixel counts (collected from images) found moderate correlation using unsupervised image segmentation (R2 0.52-0.89) and high correlation between supervised image segmentation (R2 0.91 – 0.95). Using game cameras to collect streamflow parameters was more successful on smaller streams and areas with natural sun and rain protection. One major limitation of this method is not collecting images at night, which is generally problematic during winter months. However, this method has additional benefits, such as images capturing qualitative data such as bank inundation and fish timing and presence. Using recursive feature elimination and random forest models, Chapter 3 explored which variables (climate, timing, and basin characteristics) are important in predicting low flow occurrence in rain and rain-snow dominant watersheds in Southeast Alaska. Climate and timing variables selected for all model types include Julian day, week, 7, 14, and 21-day moving precipitation average, 14-day moving temperature average, and annual snow water equivalent (SWE) maximum. Only the basin characteristics of area and max elevation were chosen for all appropriate models. Filtering for certain climate conditions and basin characteristics within a watershed improved the low flow occurrence prediction. For example, selecting watersheds with a 21-day moving precipitation average less than 0.5 mm/km2 the percent low flow correct prediction was between 38% and 84%. The results of this thesis will help support watersheds management in Southeast Alaska.

How Our Software Was Used

Dragonfly’s Deep Learning Tool was used to segment the image sets and once classified, the ROIs were exported as binary image stacks and the pixels classified as water were totaled for each image.

Author Affiliation

(1) Oregon State University