Pretty much what the subject line says. I manage data for a relatively large chunk of land mostly used for forestry. During every harvest project, a significant amount of roads are constructed or renovated. Naturally, record keeping has been sparse or non-existent, and when the specialists scout out locations for planned projects, they consistently find that the existing line feature services are inaccurate, contain roads that never could have existed, or does not include roadbeds that do.
We have fairly recent 1m LIDAR DEM coverage for most of the district, and foresters do regularly hand-digitize roads from observations on the slope raster. I'm wondering about the best method to automate or semi-automate classification of roadbeds (cells where the slope is very consistent relative to the terrain above and below the roadbed). I'm fairly new to imagery classification so I'm taking in all the ideas for training classification algorithms; I've also seen posts out there suggesting using least-cost path. I'm hesitant about the latter option; I'm not simply trying to find individual roads with known start points, but to generate predictive lines over as much of the managed area as possible.
Thanks to already having a fairly attribute-heavy feature service of the road network, I hope that the majority of the work after yielding a usable line layer would be replacing geometry on road features where the geometry is the difference between recorded road and observed road.
I'm away from work for the week but I'm hoping to get some ideas churning before I return; however that also means I can't test them out just yet. Any and all input is welcomed!
I wonder if a pre-trained model for unpaved/dirt roads as referenced in this blog article would help with the capture process? Repurposing Deep Learning Models using Transfer Learning in ArcGIS