David Nicholas Dralle , Dana Ariel Lapides , Daniella M Rempe, W Jesse Hahm

Growth and contraction of headwater stream networks determine the extent and quality of ecologically critical habitat, and open a window into the storage dynamics of catchments. A fundamental challenge is observation of the process itself: wetted channel extent is highly dynamic in space and time, with the length of wetted channel sometimes varying by orders of magnitude over the course of a single storm event in headwater catchments. To date, observational datasets are produced from boots-on-the-ground campaigns, drone imaging, or flow presence sensors, which are often laborious and limited in their spatial and temporal extents. Here, we evaluate high-resolution, multi-band satellite imagery as a means to detect wetted channel extent via machine learning methods trained using existing wetted channel extent surveys. Even where channel features are smaller than the spatial resolution of the imagery, the absence or presence of surface water may nevertheless be imprinted upon the spectral signature of an individual pixel. We leverage existing wetted channel extent surveys at two oak savanna catchments in northern California with minimal riparian canopy cover and highly dynamic wetted channel extent due to small subsurface water storage capacity and saturation overland flow. We train a random forest model on high-resolution ($\sim$5 m pixel) RapidEye satellite imagery captured contemporaneously with the existing surveys. Withheld test data indicates prediction accuracy of wet vs. dry channel extent is $>$91\%. This predictive ability is used to produce length-discharge (L-Q) relations and to calculate spatially distributed estimates of channel hyporheic flow capacity and exchange. A sharp break in hyporheic flow properties occurs at the transition from main stem channels to lower order tributaries, resulting in a stepped L-Q relationship that cannot be captured by traditionally used power law models. Remotely sensed imagery is a powerful tool for producing wetted channel maps at high spatial resolution ($\sim$10 m in this study to channels with $>$ 0.01 km$^2$ contributing area).