Confusion Matrix from Groundtruthed GPS Points

[rach_crdnl]

Hi Folks, I'm running into some issues while using the classification wizard in ArcGIS. So I'm classifying drone imagery to identify underwater grasses, and I'm training the model based on GPS points we collected in the field that have confirmed presence of the grass. I created a 70/30 training and testing split of the GPS points. I want to create a confusion matrix to evaluate the classifier, and I want to use the testing points for validation. But it doesn't seem like it is possible to create accuracy assessment points from groundtruthed points in a shapefile. Is there a way to do that? I can't randomly select points for accuracy assessment points because the underwater grass we're trying to detect can't be reliably verified by the naked eye from imagery/map. I was originally doing a confusion matrix a round-about way, but it was calculating it based on the number of points, when I actually want the confusion matrix to be based on the number of pixels. 

[DanielFox1]

Hi @rach_crdnl 

I wonder if this workflow will help in anyway 

1. Convert Groundtruthed Points to Pixel-Based Reference
To evaluate classification accuracy at the pixel level using your GPS points:
• Buffer your GPS points slightly (e.g., 1–2 meters) to create small polygons.
• Use these polygons to extract pixel values from the classified raster.
• Assign each pixel within the buffer the known class label from your GPS data.

This gives you a pixel-based ground truth raster derived from your field data.

2. Use the “Compute Accuracy For Pixel Classification” Tool
ArcGIS Pro’s Image Analyst extension includes a tool specifically for this:

Compute Accuracy For Pixel Classification (Image Analyst)—ArcGIS Pro | Documentation

• Inputs

• Your classified raster.
• Your ground truth raster (created from buffered GPS points).
• Output:
  • A confusion matrix based on pixel counts, not point counts.
  • Includes metrics like user accuracy, producer accuracy, kappa, and IoU.

Some considerations also is to ensure that your GPS points are accurately georeferenced and match the raster coordinate system

Avoid random sampling 

If your GPS points fall between pixels, consider resampling or nearest-neighbor extraction

I hope this helps