Cloud Digital Leader

The model has to tell extreme weather apart from ordinary weather, so the training set needs the full range of conditions rather than only the extreme events — that rules out the two "extreme weather data" options. It also has to predict for the whole country, so sampling evenly across all cities keeps the data representative.

Concentrating collection on at-risk cities biases the training set toward those locations and their local patterns, which degrades accuracy everywhere else and can make the model overstate risk in cities it saw little of.