Signature separability is calculated as the statistical difference between pairs of spectral signatures. You can use the Signature Separability window to monitor the quality of your training sites. Divergence is shown as both Bhattacharyya Distance and Transformed Divergence, with the Bhattacharyya Distance as the default calculation.
Both Bhattacharyya Distance and Transformed Divergence are shown as real values between zero and two. A zero indicates complete overlap between the signatures of two classes; two indicates a complete separation between the two classes. These measurements are monotonically related to classification accuracies.
Classes with invalid signatures are shown with separability values of ‑1.0. The invalid signatures represent training sites with the total number of pixels below the required minimum (N, the number of selected channels).
Some signatures with the total number of training pixels above the required minimum but below the recommended number (5*(N*N+N)) may have a numerically singular class covariance matrix. The separability values for such signatures are also shown as ‑1.0.
The invalid signatures should be corrected, either by collecting additional training pixels, or by merging them with other signatures.
To open the Signature Separability window:
© PCI Geomatics Enterprises, Inc.®, 2026. All rights reserved.