The index to the articles in this series is found here.
With a few tweaks to the phantom rain detection, suddenly things are looking much better. My two hour warning time is now 0.83, and the failed predictions are difficult cases. Examining the predictions that failed shows rain that forms directly overhead, with no warning on radar, or small rain pockets that could either pass by the city or pass overhead. They’re cases that I wouldn’t be able to predict better by looking at the same images. So, that’s a success. I can feed a set of radar images into the neural network and have it tell me whether or not to expect rain.
What’s left to do? Now we can still experiment a bit to see if we can find a way to reduce the fraction of failed predictions. Ideally we want incorrect predictions to have lower confidence than successful ones. By histogramming the (floating point) predicted values for successful and failed predictions, I hope I can come up with a confidence level. A value of less than 0.02 indicates a firm prediction of false. A value of greater than 0.98 indicates a firm prediction of true. In between, we have lower confidence warnings. There’s a cutoff at 0.5 between predicting positive or negative results. Shifting that cutoff has the effect of trading false negative for false positives, or vice versa. I’m not going to focus on this, though.
So, I’m going to try to play around with network settings now to see if I can separate the false positives from true positives, and conversely for negatives. Right now, there’s a handful of false negatives with a value less than 0.02, which indicates a firm false prediction. I’d like to train a network that minimizes those high confidence incorrect predictions.
Postings will probably slow down a bit now, as I experiment with settings and see how changes behave.
In the mean time, I’m happy with the results of this project.