HPCCprojects:Using Neighborhood-Algorithm Inversion to Test and Calibrate Landscape Evolution Models

Revision as of 17:18, 19 February 2018 by WikiSysop (talk | contribs) (Text replacement - "http://csdms.colorado.edu/wiki/" to "https://csdms.colorado.edu/wiki/")
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Using Neighborhood-Algorithm Inversion to Test and Calibrate Landscape Evolution Models

Project description

Landscape evolution models use mass transport rules to simulate the development of topography over timescales too long for humans to observe. The ability of models to reproduce various attributes of real landscapes must be tested against natural systems in which driving forces, boundary conditions, and timescales of landscape evolution can be well constrained over millennia. This project aims to test and calibrate a landscape evolution model by comparing it with a well-constrained natural experiment using a formal inversion method to obtain best-fitting parameter values.

Our case study is the Dragon's Back Pressure Ridge, a region of elevated topography parallel to the south central San Andreas Fault that serves as a natural laboratory for studying how the timing and spatial distribution of uplift affects topography. We apply an optimization procedure to identify the parameter ranges and combinations that best account for the observed topography. Direct-search inversion models can be used to convert observations from such natural systems into inferences of the processes that governed their formation through the use of repeat forward modeling. Simple inversion techniques have been used before in landscape evolution modeling, but these are imprecise and computationally expensive. In this project, we are applying a more efficient inversion technique, the Neighborhood Algorithm (NA) (Sambridge, 1999), to optimize the search for the model parameters values that are most consistent with the formation of the Dragon's Back Pressure Ridge through repeat forward modeling using CHILD.

Inversion techniques require the comparison of model results with direct observations to evaluate misfit. For our target landscape, this is done through a series of topographic metrics that include hypsometry, slope-area curves, and channel concavity. NA uses an initial Monte Carlo simulation for which misfits have been calculated to guide a second iteration of forward models. At each iteration, NA uses N-dimensional Voronoi cells to explore the parameter space and find the zones of best-fit, from which it selects new parameter values for the forward models. As it proceeds, the algorithm concentrates sampling around the cells with the best-fit models. The resulting distribution of forward models and misfits in multi-parameter space can then be analyzed to obtain probability density distributions for each parameter.

The ability of NA to provide probability distributions for parameter values gives an indication of uncertainty in each, and can be used to guide field measurements for model testing. This application of advanced inversion techniques for landscape evolution modeling is a significant step towards the use of more formal mathematical methods in geomorphology that are already applied by other disciplines in the geosciences.


The first iteration of 300 CHILD runs is taking place on the CSDMS HPCC during the late summer 2011. Further iterations (approximately 1000 more CHILD runs) will occur during the fall of 2011. The project is expected to finish by the end of the year.

Models in use

This project uses the inversion algorithm of Sambridge (1999), home-built scoring functions (to be submitted to the model repository in the future), and the landscape evolution model CHILD as the forward model.


Publications and presentations

  • Using Neighborhood-Algorithm Inversion to Test and Calibrate Landscape Evolution Models: Mariela C Perignon, Gregory E Tucker, Peter Van Der Beek, George E Hilley, Ramon Arrowsmith. AGU Fall Meeting 2011 (EP21. Quantifying Geomorphic Processes and Landscape Evolution: Linking Observations and Models).


Neighborhood Algorithm