Help using deep learning to detect larger structures in DTMs

[Cli]

Hi, I was hoping someone could give me some advice. I'm trying to use deep learning to detect drained peatlands and marsh trenches. The input raster is a hillshade of dtm.

The first problem is my training data. Upon inspecting the tiles many of them was not on any trenches. Any tip on how to train on larger structures? 

The second problem is that the output detected objects are mainly single lines (like roads) and not on trenches which looks like this:

Hope someone can help

[RonaldHaug]

Hi Cli,

I like that you are trying to locate drained peat bogs and marshes using remote sensing data. Wetlands loss greatly contributes to the disruption of the hydrologic cycle and high pressure heat domes. Good on you!

First off, you have to be able to know what drained wetlands look like on the ground, if that is even remotely possible, given the huge areas which have been drained and turned into agriculture and urban areas over the millennia.

Once you have something you can model, then you can train the program to look for areas which meet the criteria. Digital Elevation Models (DEMs), what you call DTMs (Digital Terrain Models) may provide some of the information you need. Have you considered other remote sensing data?

Try a google search for research papers which have explored the question you are trying to answer, and let us know what you find. For example, here is one I found by A Rabine at St. Mary's University of Minnesota.