deep_learning

Etude: deep learning in air pollution

Exploration of air pollution mapping and others. The very initial motivation is to learn the road-no2 relationships automatically than using buffers

• In the folder "CNN" you will find a jupyter notebook for the modeling process.

current findings in setting

*I obtained higer accuracy with dropout and batchnorm. The lowest mae is around 10.

• Averagepooling is not as steady as maxpooling. Intuitively, averagepooling makes more sense because we always aggregate road length withn a buffer.
• Batch size setting to 100 obtained better results than 32 and 50.
• The activation of the last layer is set to relu to indicate no negetive predictions and for better backprop.
• With imageGenerator it seems to obtain a more steady result.