ICM gives invalid threshold

[ashnair1]

Describe the bug

Running IRMAD with initial change mask results in an invalid threshold. Plot attached

To Reproduce

Run

changedet --algo irmad run data/LT5_19980329_sub.tif data/LT5_19980516_sub.tif --niter 50 --icm

Expected behavior
The threshold should lie between the No-change and change distributions.

Screenshots
If applicable, add screenshots to help explain your problem.

Plot of invalid threshold

Desktop (please complete the following information):

• OS: Ubuntu 18.04

[ashnair1]

With 6b41a3d, the threshold seems to be computed correctly.

However there are two problems:

1. The previous np.roots method and the current discriminant method should have given the same result. The fact that it doesn't implies there might be an error in the derivation or in translating the equation to code. Derivation can be found here.
2. The method still fails for my current test data. This line errors out with a numpy.linalg.LinAlgError: Singular matrix error. This isn't particularly informative. A better error message should be raised and the cause of this needs to be identified. It could be that the two images don't have sufficient unchanged pixels but I'm not sure.