Abstract

Signature extension is a process intended to increase the spatial-temporal range over which a set of training statistics can be used to classify data without significant loss of recognition accuracy. The goal of signature extension is to minimize the requirements for collecting ground truth and extracting training statistics, thus reducing the costs and time delays associated with those procedures. Signature extension would then help to provide timely and cost-effective classification over extensive land areas, including remote areas for which ground truth information may not be readily available.

Many current signature extension techniques are based on a transformation of training statistics to compensate for changes in sun angle, atmospheric conditions, etc., between a training area and a recognition area. Although preprocessing techniques which minimize or eliminate the need for altering training statistics are also potential solutions to the problem of signature extension, this presentation is principally concerned with those algorithms which define signature transformations based on associations between training and recognition area statistics.

ERIM has shown that since causes in nature for variations in the measured radiance from a given material are in all cases multiplicative and/or additive, an appropriate signature transformation would he both multiplicative and additive in each data channel. In principle, this signature transformation should be unique for each material since bidirectional reflectance, influenced by such factors as sun angle, wind velocity, and soil variations, is a unique attribute of each type of ground cover. However, current signature transformation algorithms concentrate, with only a few exceptions, on defining an average transformation to be applied equally to all signatures. A first cluster matching algorithm (called MASC, for multiplicative and Additive Signature Correction) was developed at ERIM to test the concept of using associations between training and recognition area cluster statistics to define an average signature transformation.

A more recent signature extension module, CROP-A (Cluster Regression Ordered on Principal-Axis), has shown evidence of making meaningful associations between training and recognition area cluster statistics, with the clusters to be matched being selected automatically by the algorithm. These associations have led to multiplicative and additive signature corrections producing classification results over recognition areas which were significantly improved relative to what would have been achieved without the signature transformation and without local training.

The manner in which a signature extension module such as CROP-A, is embedded in an overall signature extension system has been identified as an important consideration in determining its performance and value as a signature extension tool. In this regard, research is currently underway a ERIM to define an optimum signature extension system utilizing the current state of the art. Improved signature extension modules are currently undergoing development, test, and evaluation.

Partitioning (i.e., defining the limits of regions over which a signature extension technique can reasonably be applied) has been identified as another major factor controlling signature extension utility. Hence, current research is also concerned with defining the necessary factors which limit the extent of a partition.

Date of this Version

1976

Share

COinS