Accuracy optimization for high resolution object-based change detection: An example mapping regional urbanization with 1-m aerial imagery

Abstract

The utility of land-cover change data is often derived from the intersection with other information, such as riparian buffers zones or other areas of conservation concern. In order to avoid error propagation, we wanted to optimize our change maps to have very low error rates. Our accuracy optimization methods doubled the number of total change locations mapped, and also increased the area of development related mapped change by 93%. The ratio of mapped to estimated change was increased from 76.3% to 86.6%. To achieve this, we used object-based change detection to assign a probability of change for each landscape unit derived from two dates of 1 m US National Agriculture Imagery Program data. We developed a rapid assessment tool to reduce analyst review time such that thousands of locations can be reviewed per day. We reviewed all change locations with probabilities above a series of thresholds to assess commission errors and the relative cost of decreasing acceptance thresholds. The resultant change maps had only change locations verified to be changed, thus eliminating commission error. This tool facilitated efficient development of large training sets in addition to greatly reducing the effort required to manually verify all predicted change locations. The efficiency gain allowed us to review locations with less than a 50% probability of change without inflating commission errors and, thus, increased our change detection rates while eliminating both commission errors and locations that would have been omission errors among the reviewed lower probability change locations.