Discrepancy Detection

One project in the lab relates to detecting discrepancies in sensor data and then prioritizing updates based on those discrepancies and their expected impacts. The approach uses a fair amount of probability to arrive at posterior probabilities of discrepancies along with expected net value. You might find that some data is certainly wrong due to, say, obsolescence, but you might also believe that nobody will care. Other data might be less likely wrong but still questionable. If that data is particularly valuable, then you might want to spend some resources to update it.

Maps and imagery are one application. The system processes many map and imagery tile samples on a regional basis. From a map:




From imagery:




Then we compute posterior probabilities of meaningfully stale imagery, as represented here




The colored features are present on the map but not the imagery. Some of the roadwork is in progress but incomplete, and other road construction hasn't begun. If this image tile is important to you, then you might want schedule an update pretty soon.

The analysis required much more image processing and statistical work than we had originally thought, but the results are pretty good. Encouraging enough that we will probably proceed with building a production system for large-scale processing. The framework should be general enough to handle a reasonably wide range of sensor data types as well as some annotations. Let us know if you'd like early access once we get things going.