Authors: Smruti Naik*, Clark University, Emily Evenden*, Clark University, Isaack Okwaro, Clark University, Zhenhua Meng, Clark University, Robert Gilmore Pontius, Clark University
Topics: Spatial Analysis & Modeling, Geographic Information Science and Systems, Environment
Keywords: Land Change Modeling, Sensitivity Analysis, Machine Learning
Session Type: Poster
Presentation File: No File Uploaded
We performed sensitivity analysis on how TerrSet’s Land Change Modeler (LCM) simulates the transitions among three categories: Water, Marsh and Upland. LCM used transitions during 1938-1972 for calibration then extrapolated transitions during 1972-2013. We examined how the allocation of extrapolated transitions varied among 16 combinations of driver variables: Elevation, Slope, Aspect and Distance to Categories’ Edges at 1972. Validation during 1972-2013 generated four components: Misses, Hits, Wrong Hits, and False Alarms. Hits are correctly simulated transitions, while the other three components are errors. All 16 combinations of driver variables produced more errors than Hits at the pixel level. Most of the error was due to erroneous allocation as opposed to erroneous quantity of simulated transition. Aspect, Slope, and Elevation was the combination of driver variables that produced the greatest number of Hits. Elevation was the single driver variable that produced the greatest number of Hits. LCM uses a neural net to fit a relationship concerning transition potential versus the driver variable(s). The neural net fit a relationship whereby lower elevations of Marsh transition to Water while higher elevations of Marsh transition to Upland.
To access contact information login