How do we improve regression model model?


Preprocessing


For example, we may preprocess spectral data. In our example, we may use 1st spectral derivative and retrain the model. Updated model shows better fit:



perClass Mir provides R^2 and Q^2 statistics quantifying model fit in the regression plot legend. The R^2 refers to training objects and Q^2 to test objects. Our goal is to get as good test set fit as possible, meaning the highest Q^2 statistic.


Feature subset

In some problems, it may be beneficial to avoid certain parts of spectrum in regression modeling. Typically, start and end of the spectral range exhibit higher noise and may be better avoided. We may simply select desired set of bands in the spectral pannel.


perClass Mira allows to use different feature subset for the pixel classifier and different for the regressor. Both models adopt the selected bands when trained.


Limiting spatial neighborhood

In non-homogeneous objects, it may be useful to limit the object area used for building regression model. We may do that by defining "point radius". Use the Set point radius command from Regression menu. By default, the radius is 0 which means that all spectral of each object are used to build regression model. By limiting the radius to, for example 12 pixels, only the spectra inside the displayed circle will be utilized.



When point radius is adjusted, regression data set needs to be re-created with Create data set button.


Note, that the point radius is applied only in building the training and test sets. When applying the model to a new image using Apply to new image entire object is used.