MAPPING PATCHY SOIL MANTLES USING LIDAR TOPOGRAPHY IN THE RAMPART RANGE, CO

The transition from soil-mantled to bedrock-dominated surfaces records the (im)balance between soil production rates and hillslope denudation rates. Better understanding this important transition requires, in part, more detailed maps of the patchiness of soils across gradients in driving variables (e.g., climate, rock uplift, lithology, ecosystems). Prior studies show that airborne lidar topography is well suited to this mapping task because: 1. Fine-scale topography is sensitive to the emergence of bedrock, and 2. It balances tradeoffs between data resolution and coverage. However, using topographic thresholds for classifying bedrock requires high resolution ‘truth’ maps at select locations. Truth maps are used to determine optimal thresholds before applying classifiers at the landscape scale. Thus, assessing the accuracy of lidar classifiers of bedrock is paramount. Using a suite of synthetic scenarios, we show how pixel-based accuracy metrics are expected to vary as a function of bedrock fraction. Specifically, we show why Matthews Correlation Coefficient should be favored over other commonly used metrics like precision, recall, and F1-score. While it is well known that Matthews Correlation Coefficient is more robust to imbalanced data, this work shows how the limitations of other metrics are exacerbated when the bedrock classifier is expected to perform across large gradients in soil fraction. As such, we use Matthews Correlation Coefficient to build and test two bedrock classifiers in the Rampart Range, CO. This setting has large gradients in fractional soil cover as a function of elevation and aspect. Bedrock frequently emerges as tors, or isolated bedrock outcrops within a soil mantle. We find that while slope-thresholds work reasonably well in this setting, they often fail to identify the tops of tors because steep side-slopes are connected by low-slope tops. Our new classifier instead uses slope-thresholds to identify and label individual features. Feature boundaries are then derived using the concave hull polygon that surrounds labeled features. When both methods are optimized with respect to Matthews Correlation Coefficient, our new classifier outperforms the slope-threshold proxy. We attribute the improvement in performance to be due to the high frequency of tors in this setting.