GeoAI Tool-Extract Features Using AI Models
I trying to use this guide Use the model—ArcGIS pretrained models | Documentation to extract building footprints in Africa using the pretrained model Building Footprint Extraction - Africa - Overview (arcgis.com).
When select the dropdown to select the pretrained model, you get a certain set of these pretrained models shown in the screenshot but not the one I want. Can I add the Africa one or others?
I believe so. If you download the Building Footprint Extraction - Africa - Overview (arcgis.com) model locally on disk, then in the Models Dropdown parameter "additional models" - navigate to where the downloaded model is located and set the Postprocessing workflow to "polygon regularization" This should work for you.
I learned a little more about the Extract Features Using AI Models. Regardless of where you access the Extract Features Using AI Models; from GeoAI tools or Image Analyst tools, it is the same. I discovered that if you select the Pretrained Models first without selecting an input raster, you get the long list of pretrained models. I've learned that the pre-trained model requirements for the building extraction as described in the description (in this case as the 8-bit, 3-band high-resolution (10–40 cm) is very important. These characteristics determine which pre-trained models will appear in the dropdown. I was working with a 0.5 pan-sharpened Worldview 2. When I added a 0.3 m version of image, the pretrained Africa Building Extraction model appeared in the dropdown. Below is a screen of the image. The yellow polygons are the extracted features using the Extract Features Using AI Models with the WV2 image saved as 0.3 meters. The red-filled polygons are the DL results using the original pansharpened image at 0.5meters and adding the Model under "Additional Models."
The Segment Anything Model and 0.5m WV2 pansharpened is shown in the second screen shot with pink buildings produces much better results, but abutting buildings are not always identified or there is duplication (a polygon within a polygon), and trees are identified at times.
WV2 panchromatic image (original resolution of 0.5m rescaled to 0.3m). Red filled polygons extracted using the 0.5 meter image, yellow outlined polygons used the 0.3m image. Used the Extract Features Using Air Models and the pretrained AGO Building Extraction for Africa with default parameters.
WV2 0.5 meter resolution image Segment Anything Model (results shown in pink)
Zoomed in SAM results
