Using Unsupervised Learning to Determine Dominant Physical Features that Control the Land Surface Temperature Variability over the CONUS Mountain West
Abstract
Watershed ecohydrological function is influenced by various physical features: vegetation, elevation, soils, and climatology, among others. In the last decade, remote sensing and GIS capabilities have reached unprecedented levels, providing high-resolution images of these features. Generally, the observed images can serve as proxies for the heterogeneous processes of the water and energy cycles determining watershed function. Traditionally, identifying functionally similar units has been used to understand the covaried heterogeneity between the physical environment and the processes it influences. The methods have evolved from approaches using isolated axes of variation to clustering techniques that summarize the covariance between several axes. Despite the advances, there remains a need for a strategy that allows identifying the most important axes of variation for specific ecohydrological processes. Considering the multi-scale nature of heterogeneity, this strategy is expected to depend on the features' and processes' spatial aggregation.
This presentation aims to identify the critical features, determining the spatial heterogeneity of remotely sensed summer daytime mean Land Surface Temperature (LST) over the Mountain West region of the Contiguous U.S. We use the GOES-16 LST product to compute mean values and assess the functional similarity of the watersheds. The physical features include elevation, soil properties, tree cover fraction, climatology of precipitation, air temperature, and incident solar radiation. We cluster watersheds with an extensive set of weights (i.e., importances) independently assigned to each feature. From this set, we select the optimal weights as those that minimize the variability of LST within each cluster and maximize it across different groups. Additionally, we test if using the feature importance obtained from a supervised learning approach (i.e., random forest regressor) as weights within the clustering leads to similar results to those obtained using the extensive-clustering optimization method. In parallel, we evaluate the impact of different scales and sub-regions by defining watersheds at different scales in the USGS Hydrologic Units Codes (HUC) system: from HUC8 to HUC12. We perform each analysis over every HUC4 sub-region within the domain.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H35C..02T