search

wri / UrbanLandUse

Characterizing urban land use with machine learning
29 stars 10 forks source link

6-category models: training, mapping, scoring #28

Open pkerins opened 5 years ago

pkerins commented 5 years ago

There appear to be some areas in the code base that still have a hard-coded assumption of 3 categories for non-road models. For example, I believe that when raw prediction values are saved during mapping, they are stored in a 3-band tif, without regard for the model typology. All parts of the code base related to training, mapping, and scoring models need to be checked for this type of hard-coding, and instead use parameters or some inference (eg model output nodes) to account for the possibility of non-trinary classification.

pkerins commented 4 years ago

To start, let's review the 6-category mapping workflow that is already being used for Mexican cities, and see where manual changes were required.

pkerins commented 4 years ago

Want to bring special attention to correcting model scoring, whether at creation or application. This is heavily hard-coded. Want to make this robust enough to handle up to 7-category models. Practically speaking, it will be 3-category, 6-category, or 2-category (roads) models. Will make something ugly but robust--one size fits all--and close the book there. An early priority for next development phase will be a fresh recordkeeping paradigm.