The purpose of this paper is to analyze the sensitivity of Geographic Information System outputs to errors in inputs derived from Remotely Sensed Data (RSD). The attention is restricted to outputs of suitability models with "per cell" decisions with gridded Geographic Data Bases (GDB) whose cells are larger than the RSD pixels. The procedure for merging RSD into such GDB's involves classification, registration and aggregation. The first two steps introduce errors at individual pixels and the last step tends to compensate for such errors.

The classification and registration errors are treated independently for the purposes of analysis. Under certain simplifying assumptions, the probability of misaggregation (that is, wrongly assigning a cell after aggregation) is expressed in terms of the probability of misclassification. A Monte Carlo simulation has been performed to show the effects of misregistration on the cell assignments.

Experiments were performed with a data base covering the Harrisburg, Pennsylvania, area. Landsat data covering the same area were classified and registered to the data base.

A baseline data set was prepared as accurately as possible. Perturbations were introduced in the form of (i) classification errors at locations of low confidence in the multispectral classification and (ii) registration errors by selection of subsets of ground control points from those used for the baseline. The errors before and after aggregation and after using the aggregated data in a suitability model were determined using pixel by pixel comparison. For this experiment, combinations of the classification and registration errors were also used.

It is found that approximately 50% reduction in error occurs due to aggregation when 25 pixels of RSD are used per cell in the GDB. Further reductions in error occur during the modelling process depending on the percentage of the total number of cells affected by RSD.

Date of this Version