Presentation on theme: "Best practices to evaluate land change models Robert Gilmore Pontius Jr Clark University, USA 1."— Presentation transcript:
Best practices to evaluate land change models Robert Gilmore Pontius Jr Clark University, USA 1
Researchers should consider … 1.uncertainties that can derive from questionable data quality, unclear boundary conditions, inappropriate model structure, and non-stationarity processes. 2.sensitivity analysis that examines the variation in model output due to specific amounts of variation in model input, parameter values, or structure. 3.pattern validation that compares model outputs with observed outcomes showing misses, hits, wrong hits, false alarms, and correct rejections. 4.structural validation to consider the consistency between real world processes and the processes that the model portrays. 2
Three Base Maps 3
Map of Accuracy at 32 m Resolution 4
Bar of Accuracy at 32 m Resolution 5 Simulated Change Reference Change
Figure of Merit is 21% 6 Numerator Denominator
7 Pontius and Si Figure 1 We want to know whether the black patches of gain of disturbance in the upper image tend to occur on the higher ranking darker pixels in the lower image. Red is eliminated because red is not a candidate for gain of disturbance.
8 Pontius and Si Figure 2 The TOC plot above shows information to show all four entries in the 2-by-2 matrix for each point. The TOC plot avoids ratios. The ROC plot fails to show sufficient information to produce the entries in the 2-by-2 matrix.
9 Pontius and Si Figure 3 The TOC plot is a transformation of the ROC but the TOC has one more bit of information, which is the reference prevalence. So TOC has all the properties of ROC and more. TOC can compute the Area Under the ROC curve, which is a popular measure. If you have an ROC from a clinical trial, then an estimate of the prevalence will allow you to construct an estimate of the TOC.