Realtime Anomaly Detection and Subpixel Object Detection in Hyperspectral Imagery

During search and rescue (S&R) operations, large areas must be visually scanned. Operator's fatigue and attention span limits the effectiveness of such visual searches. Hyperspectral sensors operating in the VNIR portion of the spectrum may provide potential alternative. These sensors may be reasonably priced for use by organizations such as the Civil Air Patrol, Coast Guard, etc., and may provide real-time cueing system to the S&R pilot. The disadvantage of a VNIR hyperspectral sensor is that it does not provide all weather and night operations. SAR systems do allow such operations, but they are much more expensive.

OKSI demonstrated the potential use of hyperspectral data for S&R applications using HYDICE data collected over a target field with panels of various sizes, different materials and paints. Some targets were as small as 1/10 of the pixel size (subpixel targets).

The subpixel anomaly detection algorithm uses a 3×3 sliding mask that compares the spectral signature within each pixel to its neighboring pixels. A ranking scheme is used to select pixels that exhibit the most anomalous features.

Each pixel is examined 9 times with a 3×3 sliding mask

The technique is very sensitive, and as shown in the figure below can detect not only the panels in the field but some bushes at the edge of the field. These natural objects are then sorted out from the manmade objects by use of a proper clustering algorithm that operates only on the "suspect" pixels.

The performance of the detection algorithm has not deteriorated when only the first 70 (out of 210) HYDICE bands are used. These are the bands that correspond to the VNIR portion of the signature, indicating that a sensor based on a CCD FPA can be used for this application.

Return to OKSI's homepage
Go to TechExpo Homepage.