Total Monthly Rain forecast focused as characteristics vs magnitude classification with CNN

Hello: Excuse me if this is a very basic query at the level of this forum (*):

8/9 years ago, I assembled Excel macros and Matlab scripts to build data preprocesses and Canonical Correlations (linear) to Forecast rainfall totals (mm) 1, 2 and 3 months ahead in Neuquén/Limay basins, southern Argentina . The target samples were 40 years old.

As Inputs, it used 3 to 4 oceanic and atmospheric variables from up to 2 previous months, from NCEP/NCAR NOAA Reanalysis matrices selected and analyzed with PCA (EOF).

I did not use the chronological sequences as a time series. One correlation (I would say "static") resulted from each month, for each of the three times ahead. Thas is, 40 observations/correlation. I compared results with cross-validation, leaving one year out, each time.

In order to try to improve the results of that CCA with non-linear methods and so little data extension and to avoid overfitting, ¿could you approach training and testing using CNN, using target precipitation quintiles or deciles for each forward time as outputs?

That is, as if it were a classification problem instead of obtaining an absolute value in (mm).

Thank you very much!!! Ginés.