Watershed Monitoring Network Optimization Through Co-Design: Bedrock-to-Canopy Characterization and Ecohydrological Modeling
Abstract
In this study, we have developed a machine learning (ML)-enabled framework for selecting hillslope-scale ecohydrological experimental sites in snow-dominated mountainous watersheds and for optimizing the sensor network design for snow-soil-plant interactions. We take advantage of satellite and airborne remote sensing technologies (LiDAR, hyperspectral, snow LiDAR, electromagnetic, Landsat), as well as integrated ecohydrological models, providing various spatial layers for bedrock-to-canopy properties and watershed functions, such as evapotranspiration (ET), water table dynamics, and drought sensitivity. In this framework, we first applied unsupervised clustering to the hillslope-averaged spatial data layers, and mapped the clusters in space as zonation that captures the watershed-scale heterogeneity of bedrock-to-canopy properties. We merged the results from different clustering methods (Gaussian mixture models, K-means, hierarchical clustering) for selecting the most representative hillslopes in each zone. In addition, to account for the within-hillslope variability, we developed an algorithm to optimize the number and placement of sensor packages of soil moisture, precipitation and snow. We combined the Gaussian mixture models and Gaussian process models to capture the co-variability of various spatial data layers and to improve the spatiotemporal interpolation. We demonstrate this framework, using the datasets collected in the East River Watershed, Crested Butte, Colorado, USA. We found that the bedrock-to-canopy properties were significantly correlated with each other so that zonation could reduce the dimension of those properties effectively to a small number of zones for capturing their spatial heterogeneity. In addition, our approach documented the importance of high-elevation alpine regions for ecohydrological functions (such as water table dynamics, soil moisture and ET), which are often neglected due to the access difficulties. Through using ML to explicitly bridge information derived from on the ground" observations, remote sensing data, and numerical models, we aim to quantify fundamental scientific linkages among interacting processes in the watershed and to improve the predictability of watershed functions.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H31B..08W