Training Deep Learning Models with a single band (grayscale) Raster

NinaHarvey1 · ‎11-22-2021

I'm trying to follow the Label Objects for Deep Learning Workflow but every time I try to Export Training Data, my Pro (2.8.0) crashes.

I am signed in to my esri account and here are my export parameters:

Export Training Data
Output Folder: DeepLearning\Trained_Data
Mask Polygon Features: DeepLearning\DeepLearning.gdb\TS_Attempt1 (.shp)
Image Format: TIFF Format
Tile Size X & Y: 256
Stride X & Y: 128
Rotation Angle: 15
Reference System: Map Space
Output No Feature Tiles: False
Meta Data Format: PASCAL Visual Object Classes

Also I am finding very little information online about how to follow any Deep Learning classification workflows with single band, monochromatic, raster datasets (I am working with historical air photos).

I think the Training Samples Manager workflow will also be limited as I read somewhere that the Classification Tools menu is unavailable if the highlighted image is not a multiband image.

Can anyone point me to a workflow or resource that talks about Deep Learning Models with single band images?

DrVSSKiran · ‎11-22-2021

It would be great if you could share your system hardware specification.

If your system GPU is less than 4 or 6 GB then try to make the tile size is 128.

Thanks & Regards

ABishop · ‎11-22-2021

Hello Nina,

Since you mentioned you were working with single band, monochromatic, raster datasets that are historic in nature, I found two links that may be of interest. The first link discusses the Deep Learning Model and compatible data formats. The second link talks about a deep learning model which can assist you with colorizing historic black and white imagery which may enable you to execute the primary solution.

https://pro.arcgis.com/en/pro-app/latest/tool-reference/image-analyst/overview-of-the-deep-learning-...

https://www.esri.com/arcgis-blog/products/arcgis-pro/design-planning/colorizing-historic-black-and-w...

Amanda Bishop, GISP