using SOM for remote standoff detection
Use case

Let’s introduce here an active standoff detection system of aerosol based on the principles of infrared polarized elastic scattering and the 4 x 4 Mueller matrix. SAID MORE SIMPLY : let’s try to identify what kind of aerosol we can mesure in a space by using a scattering system measurement.

If you don’t like technological explanations, skip the following : sensor generates all 16 matrix elements instantaneously in groups of 8 and modulo-2 laser beam interrogation wavelengths. One laser beam is tuned to a peak in molecular absorption of a detection target, i.e. a chemical-biological (CB) aerosol or surface contaminate warfare agent, while the alternate beam is detuned to the tail of that absorption band. Without entering in details, this allows to define a “fingerprint” identifier. Roughly speaking, Mueller matrix from classical electromagnetic field theory is defined as a transformation of Stokes vectors between incident and backscattered electromagnetic fields.






By injecting labelled material in aerosol chamber and storing a bunch of data (time stamp, mean and standard deviation of Mueller Matrix elements), we get a data set that we randomly split in a training and validation sets (we do not use here n-fold training process).

Then a hybrid framework containing Kohonen Self-Organizing Maps (SOM) and feed-forward neural networks identifies and classifies those CB analytes in the training set. Map below provides a view of the clustering on the training set:




Relevance of the system is then assessed by applying the SOM on the validation set:




Indeed by superimposition one can see that the SOM identifies and classifies in a perfect way CB analytes, as well offering insight into the likely structure and toxicity of compounds which are not in the Mueller knowledge repository!