ArcGIS Pro

Deep Learning with ArcGIS Pro Part 3: QA/QC Extracted Features

In parts one and two of this blog series, you learned how to prepare your environment for deep learning and best practices for using Esri’s pre-trained deep learning models. Now that you have run a model, you are ready to clean up your extracted features by removing objects misidentified as buildings and regularizing your final building polygons.

Two notes before we get into the tools:

  1. In this blog post we will focus on the output features from the Building Footprint Extraction model, but this workflow can also be adapted for other feature extraction model outputs.
  2. We will be going over some of the most useful tools that are available to refine your detected buildings output, but note that there are multiple workflows for editing features so this is not an exhaustive list.

Combine Multiple Outputs

In some workflows, you will run the Detect Objects Using Deep Learning geoprocessing tool more than once and need to combine multiple output layers to get your final building polygons. This can be accomplished with the Merge and Dissolve tools as detailed below. If you only have one output from the Detect Objects tool, you can skip these steps and proceed to clipping with parcel data.

Some buildings were likely identified in both your output results. In order to avoid duplicate features, we will dissolve the boundaries between overlapping building polygons to form one polygon for the building.

Split up Multipart Features

After completing the steps above for merging and dissolving together multiple building layers, you may notice that in some cases multiple polygons are being grouped together as one building. To ensure that we get each detected building as a single polygon feature, we will run the Multipart to Singlepart tool.

Clip to Parcel Data

If you have land parcel polygon data, you can use it to refine your results by removing any “buildings” that we’re identifies by the model but fall outside of the known parcel boundaries. Assuming the parcel data is up to date, buildings outside these bounds are likely false identificcations. If you do not have parcel data, you can skip this step.

Remove Misidentified Objects

As described earlier in this process, the deep learning algorithm is never going to output a perfect result. Therefore some smaller features such as cars or piles of construction materials can be identified as buildings. We can remove these non-building polygons based on their area, as their footprints will be often be significantly smaller than a building footprint. In this case we are deleting features smaller than 50 square meters, but this value should be adapted based on the typical size of the object you are looking to detect.

Irregularized Buildings Footprints
Irregularized Buildings Footprints

Regularize Buildings Footprints

As you can see in the image above, the shapes of our current building footprints are irregular, lacking the straight sides and right angles typical of buildings. ArcGIS Pro has a geoprocessing tool specifically designed to clean up these outputs and give you more realistic building footprint polygons:

For an additional challenge, you can calculate the typical ratio of length to width of a building in your image, and use this metric to further identify polygons that are not likely to be buildings.

Automate the QA/QC process

For those who enjoy working with Python, the following is a code snippet that will do the above job for you:

About the authors

Rami is a Solution Engineer on the National Government team supporting nonprofit global organizations and land administration teams out of the Rotterdam office. He has over 5 years of GIS experience and has been working with Esri since 2016 previously as a Platform Configuration Engineer with Professional Services out of the Dubai office. He has a degree in Landscape Architecture from the American University of Beirut.


Kate is a Senior Solution Engineer on Esri's National Government team. Based in New York City, she has a background in remote sensing and environmental science. Kate currently supports National Statistics Offices globally, helping them modernize their census and statistics operations using GIS.

Notify of
Inline Feedbacks
View all comments

Next Article

Getting to Know the Geoprocessing Synthetic Aperture Radar Toolset

Read this article