I think that the coordinates caused the model to memorize the patterns of the cities, but when we try to predict on the testing set, it doesn't work because there are different cities.

Given that the model will be applied to other locations at inference time, it generally doesn't make sense to train with any location based features even though the data curated seems to encourage it.

On the other hand there's isn't that much pollutant data to usefully train a model to predict pm2_5 concentrations solely relying on pollutant features: so training with latitude and longitude based features is what yields better scores for me.