Aerial Imagery for Deep Learning Classification

[gargarcia]

I am using Classify Pixels Using Deep Learning with one of Esri's pre trained models for land classification on high resolution aerial imagery I have. Before I got the most recent aerials I tested the process using 2022 NAIP imagery and got great results. My more recent aerial imagery is noticeably less vibrant with no leaves and brown grass. When I use the same methods on this imagery I get a lot of misidentified barren land.

Is there any pre processing I can do to improve the input data before using this tool? I messed around with the raster layer image enhancement sliders, but those don't actually change the data right? If there is nothing I can do to the imagery is the only other option to train my own model on this imagery? I am not really familiar with more traditional classification techniques if any.

[RonaldHaug · Accepted solution]

Hi @gargarcia,

There is a tutorial for this which may help you.

https://www.esri.com/arcgis-blog/products/arcgis-pro/imagery/fine-tune-a-pretrained-deep-learning-mo...

Check it out and let us know if it worked for you.

[gargarcia]

I did not know it was possible to fine tune one of Esri's models. I assumed you had to start from scratch. This does look like it will help for different imagery. Thanks!