Using imperfect data to validate a model of land change
This paper presents a general technique to validate models that predict changes in landscapes between two points in time. The methodology provides useful information about how the reproducibility of both calibration and validation maps influence the validation process. The reproducibility of a map is the expected agreement between a map that contains error and a theoretically remade map that has the same amount of error distributed independently. Matrix algebra is used to compare three maps: 1) the reference calibration map, 2) the reference validation map and 3) the model's predicted map. This comparison is conducted to validate a model of land change while considering: 1) the reproducibility of the reference calibration map, 2) the reproducibility of the reference validation map and 3) the predictive power of the model. We illustrate this technique by applying it to a study area in Central Massachusetts using the Geomod land change model. Geomod predicts land-use change from 1971 to 1999 among four categories: Built, Forest, Water and Other. Results show that the reproducibility of the calibration map of 1971 must be more than 71% in order for the predicted change in the predicted land-use map of 1999 to be measurable, and that the reproducibility of the validation map of 1999 must be more than 76% in order to detect any signal of land change in the reference map of 1999. If the reproducibility falls below these thresholds, then the model's performance should not be assessed with respect to the imperfect data.
American Society for Photogrammetry and Remote Sensing - Annual Conference 2005 - Geospatial Goes Global: From Your Neighborhood to the Whole Planet
Petrova, Silvia and Pontius, Robert Gilmore, "Using imperfect data to validate a model of land change" (2005). Geography. 780.