Seeking guidance: Achieving desired plot-level segmentation with MaskRCNN in ArcGIS Pro

[Kwakuopokuware401]

Hi

I hope this message finds you well. I am reaching out to you today to seek your guidance and expertise regarding a technical challenge I am currently facing with plot-level segmentation using MaskRCNN in ArcGIS Pro. I have recently been working on a project that involves training a model for plot-level segmentation. I have successfully prepared my training data using the ArcGIS training tool for deep learning and exported it with the appropriate metadata information. However, upon testing my trained model, I encountered an issue where the results are displayed in small chips, similar to the training labels, rather than the desired plot-level segmentation. My primary objective is to obtain segmentation results that align with the plots I have trained the model on. I would greatly appreciate any assistance or clarification you could provide to help me overcome this hurdle and achieve the desired output.

[PavanYadav]

@Kwakuopokuware401 

I'm not 100% sure, but one possible reason could be the tile size of your training data. You can try these to see if you get better results:

• Use a tile size that is 3 times larger than the average size of your object of interest.
• Try "Output No Feature Tiles" (unchecked) - only image chips that capture training samples will be exported. This is the default.
• Experiment with the "Minimum Polygon Overlap Ratio" parameter.

In the Export Training Data For Deep Learning (Image Analyst) tool within ArcGIS Pro, also make sure that the chip size is not smaller than the tile size in the Train Deep Learning Model (Image Analyst) tool.

You might find these two blogs useful:

• Tips for labeling images for object detection models (https://www.esri.com/arcgis-blog/products/arcgis-pro/geoai/tips-for-labeling-images-for-object-detec...)
• Tips for training data preparation for object detection models (https://www.esri.com/arcgis-blog/products/arcgis-pro/geoai/tips-for-training-data-preparation-for-ob...)

FYI, we have a dedicated board for ArcGIS Image Analyst questions. There, you may get answers sooner on imagery-related workflows.

Cheers!
Pavan

Pavan Yadav | Product Engineer - Raster/Imagery team

THE SCIENCE OF WHERE ®