Search

Substrate facies monitoring is critical for the understanding of fluvial geomorphologic and ecohydraulic patterns and processes. However, direct substrate measurement is time-consuming and subjected to data sparsity because of small sample, size, and limited data collections within an area of interest, which make it difficult to capture facies patterns. Most new experimental studies focus on mapping substrate based on median grain size of a specific grain size class using automatic or semiautomatic photosieving techniques. This study aimed to develop and apply a method to accurately predict size-mixture facies patterns on exposed riverbeds with minimal ground truth plots (100) using airborne lidar and machine learning. The selected testbed river was a 37.5-km stretch of the regulated lower Yuba River in California, USA, mapped at sub-meter resolution in 2017. First, we designed a grid-by-point grain size sampling method and binned grain sizes into representative mixtures, such as fine or large gravel, to assign subaerial facies labels. Second, we classified facies based on a multivariate cluster analysis. Third, we generated 15 lidar-derived topographic and spectral predictors. Six distinct size-mixture facies were identified from field data and a seventh, pure sand facies, from UAV data. A random forest predictive model with an 86% 10-fold cross-validation accuracy was applied to produce a facies map at the 1.54 m pixel scale. The detrended elevation was identified as the most important variable for predicting facies spatial patterning, followed by baseflow, wetted area proximity, and green lidar intensity. We conclude that machine learning combined with intensity lidar data is highly effective for distinguishing mixed classes of substrates. Ultimately, the new substrate mixture-binning approach also provides novel insights into the arrangement of river sediment facies patterns.

Substrate facies monitoring is critical for the understanding of fluvial geomorphologic and ecohydraulic patterns and processes. However, direct substrate measurement is time-consuming and subjected to data sparsity because of small sample, size, and limited data collections within an area of interest, which make it difficult to capture facies patterns. Most new experimental studies focus on mapping substrate based on median grain size of a specific grain size class using automatic or semiautomatic photosieving techniques. This study aimed to develop and apply a method to accurately predict size-mixture facies patterns on exposed riverbeds with minimal ground truth plots (100) using airborne lidar and machine learning. The selected testbed river was a 37.5-km stretch of the regulated lower Yuba River in California, USA, mapped at sub-meter resolution in 2017. First, we designed a grid-by-point grain size sampling method and binned grain sizes into representative mixtures, such as fine or large gravel, to assign subaerial facies labels. Second, we classified facies based on a multivariate cluster analysis. Third, we generated 15 lidar-derived topographic and spectral predictors. Six distinct size-mixture facies were identified from field data and a seventh, pure sand facies, from UAV data. A random forest predictive model with an 86% 10-fold cross-validation accuracy was applied to produce a facies map at the 1.54 m pixel scale. The detrended elevation was identified as the most important variable for predicting facies spatial patterning, followed by baseflow, wetted area proximity, and green lidar intensity. We conclude that machine learning combined with intensity lidar data is highly effective for distinguishing mixed classes of substrates. Ultimately, the new substrate mixture-binning approach also provides novel insights into the arrangement of river sediment facies patterns. [Display omitted] • Cluster analysis of field-plot grain-size data yields mixed-size sediment facies to train machine learning. • A supervised random forest machine learning model accurately predicts subaerial facies using only airbone lidar. • The top lidar predictor is detrended elevation. [ABSTRACT FROM AUTHOR]