How to Train a Deep Learning Model with Polyline Data?

[DavidSolari]

I'm exploring a process where I take polyline street features and the imagery they were HUD-ed from to train a model to spit out polylines from new imagery. I've looked over the Export Training Data For Deep Learning and Use Training Samples Manager pages and at first glance it looks like the tools are designed around extracting points & areas, not linear features. Is there a recommended workflow for training an Image to Polyline model in ArcGIS? If not, has anyone had success exporting polyline data into a format a ML framework can process?

[DanPatterson]

there are a few references around

Repurposing Deep Learning Models using Transfer Learning in ArcGIS

[DavidSolari]

Thanks Dan! There's some good info on how to prep the Export tool and the suggestion to use the Multi-Task Road Extractor model type. Unfortunately this has the usual blog post issue where many inputs aren't explained and it also neglects to mention how to prepare the polylines. Are these long features that are diced up along the training chip boundaries or were individual features drawn over individual chips? Will the tool dice my inputs for me?

[DanPatterson]

it looks like the polylines are used at the training step to identify where roads exist

[DavidSolari · Accepted solution]

I ran this through support and buried in the links they sent is this fantastic tutorial: link. It's for polygon classification and doesn't use the road detection model but I think the same general process should transfer over. It's a shame all this detail is buried here, rewriting this for the standard Pro docs could cover the end-to-end DNN modelling process fantastically.