Digital landform mapping and soil-landform relationships in the North Cascades National Park, Washington

Philip Harrison Roberts

Digital soil mapping meets current demands for soils data and increases the opportunity for scientifically based management of public resources. In this thesis I employed geospatial data and geographic information systems to characterize soils, landforms and soil-landform relationships in the rugged, mountainous terrain of Thunder Creek Watershed (30,000 ha) in the North Cascades National Park (48° 30` North, 121° West). I described and classified plants, soils and landforms at over 400 spatially referenced locations throughout the study area. I used field observations, a 10 m digital elevation model and inductive classification methods, including decision trees and random forest machine learning, to produce landform maps with a 2/3 to 1/3 split between calibration data and validation data. I obtained an expert, National Park Service landform map created from aerial photograph interpretation, topographic maps, and field observations for the evaluation of automated mapping methods. Automated and expert methods were compared with field observations. Field observations of landforms correlated best with the expert map (kappa = 0.59 and overall accuracy = 70 %). Evaluating automated approaches, the random forest classification (kappa = 0.44 and overall accuracy = 59 %) performed better than the decision tree model (kappa = 0.37 and overall accuracy = 53 %). Resulting statistical models were applied to map the entire watershed. Observations of landforms were compared with soil properties. Graphical representations of categorical soil variables show strong relationships with landscape stability and profile development. Older landforms support Spodosols and Andisols while younger, active surfaces support Entisols and Inceptisols. These trends are evident when comparing podsolization, tephra distribution, and presence of redoxomorphic features to landform classes. Continuous soil variables were analyzed with generalized least squares regression. Regression models provided poor predictions of soil attributes, questioning traditional beliefs regarding soil landform relationships. My results show promise for digitally mapping landforms in mountainous terrain. My results also suggest landforms may be less important for soil mapping. Advantages to these methods are by using an inductive, empirical approach one gains knowledge of the landscape from the data directly, hopefully proving more transferable among other steep, mountainous landscapes.

Digital landform mapping and soil-landform relationships in the North Cascades National Park, Washington

Files and links (1)

Abstract

Metrics

Details