Imagery in Action

Exercise
Map building footprints with deep
learning
Section 5 Exercise 2
April 6, 2022
 Imagery in Action

Map building footprints with deep learning

Time to complete
30 minutes

Introduction
Imagery can contain a wealth of information, from the number of buildings in a city to the
health of vegetation. However, extracting this type of data from an image is more complicated
than working with vector datasets. Historically, you would need to manually digitize each
feature that you wished to extract from an image. This manual process could involve
traditional image classification and could take weeks or years, depending on the size of the
image. Deep learning automates the process of extracting features from imagery. However,
one of the primary difficulties with using deep learning has been the time and effort necessary
to train deep learning models to analyze the data. To make this process more efficient, Esri
packaged several pretrained deep learning models for various uses. These deep learning
packages allow you to use the power of deep learning without training a model. You can use
ArcGIS Image for ArcGIS Online to analyze your data with deep learning models from ArcGIS
Living Atlas of the World and extract the desired information for your work.

Exercise scenario
You are a GIS analyst for Henderson County in Texas, exploring ways to update the existing
data in your county GIS. You have been provided an imagery layer from the National
Agriculture Imagery Program (NAIP), which acquires aerial imagery during the agricultural
growing seasons in the continental United States. With this NAIP imagery layer, you will use a
deep learning package from Esri to begin extracting building footprints.
For more information about NAIP imagery in ArcGIS, see the NAIP Imagery group monitored
by Esri.
Note: This exercise provides a simplified starting point for what is often a much more in-depth
workflow. The data and parameters used in this exercise are not intended to be a
comprehensive list of what might be required for this type of analysis.

Step 1: Open an imagery layer in a web map

In this step, you will add an imagery layer to a web map.

Copyright © 2022 Esri. All rights reserved. 1

 Imagery in Action

a If necessary, restore your web browser.

Note: If you closed your browser, browse to www.arcgis.com, sign in with your MOOC
credentials, and then click the Map tab.

b In the upper-right corner, click New Map and choose Create New Map.

c On the toolbar above the Contents pane, click Add and choose Search For Layers.

d In the Search pane, click the My Content down arrow and choose ArcGIS Online.

e In the Search For Layers field, type AthensTX_2018 and press Enter.
An AthensTX_2018 imagery layer by EsriTrainingSvc will appear in the search results.

f In the lower-right corner of the AthensTX_2018 imagery layer by EsriTrainingSvc, click the
Add button .

The imagery layer appears in the web map and is rendered with the Natural Color band
combination.
Your manager recommended that you map building footprints in a certain area, so you will
add a focus area layer to the map.

g In the Search pane, type building footprint focus and press Enter.

h In the results, locate the Building Footprint Focus Area feature layer by EsriTrainingSvc,
and then add it to the web map.

Copyright © 2022 Esri. All rights reserved. 2

 Imagery in Action

i In the upper-left corner of the Search pane, click the Back arrow , and then click Content
to view the Contents pane.

j In the Contents pane, point to or click the Building Footprint Focus Area layer, click the
More Options button , and choose Zoom To.

k Click Save and choose Save As.

Copyright © 2022 Esri. All rights reserved. 3

 Imagery in Action

l In the Save Map dialog box, specify the following parameters:

• Title: Building Extraction <your initials>

• Tags: Imagery in Action, Esri, Texas
• Summary: Web map for building extraction in Athens, Texas.

m Click Save Map.

You are now ready to begin the process of extracting the building footprints.

Step 2: Extract building footprints

In this step, you will use deep learning to extract the building footprints.

a On the toolbar, click Analysis to open the Perform Analysis pane.

b Click Raster Analysis and expand Deep Learning.

c Choose the Detect Objects Using Deep Learning tool.

d In the upper-right corner of the pane, click the Analysis Environments button to open
the Analysis Environments settings.

Copyright © 2022 Esri. All rights reserved. 4

 Imagery in Action

e In the Analysis Environments dialog box, specify the following parameters:

• Extent: Layer Building Footprint Focus Area

• Mask: Layer Building Footprint Focus Area

f Leave the remaining defaults and click Apply to apply the Analysis Environments
parameters.

g In the Detect Objects Using Deep Learning tool pane, confirm that Choose Image Used
To Detect Objects is set to AthensTX_2018.

h For Choose Deep Learning Model Used To Detect Objects, choose Choose Deep
Learning Model.

i In the Choose Deep Learning Model window, click the My Content down arrow and
choose Living Atlas.

j In the results, locate the Building Footprint Extraction - USA deep learning package and,
in the lower-right corner of the result, click Select.
Note: There will be a slight delay when you return to the tool interface as the model
arguments are read.

k Confirm that the deep learning model arguments populate.

Copyright © 2022 Esri. All rights reserved. 5

 Imagery in Action

The deep learning model arguments are additional parameters that can be adjusted as
needed to improve your analysis. They are populated by the tool from reading the Python
module on the raster analysis server. Each argument is specified by the deep learning
package that you choose. These arguments in the deep learning model are the ones used in
the Building Footprint Extraction - USA package.
Notice that the confidence threshold is currently set to 0.9. This threshold means that if the
model is at least 90 percent certain that a feature is a building, it will include it in the output.
You will run the model with the default threshold and then determine if it needs to be
adjusted lower when you review the output.
For more information about deep learning models, see ArcGIS Enterprise Help: Deep
learning in Raster Analysis: Esri model definition file.

l For Result Layer Name, type AthensTestDeepLearning_<your initials>.

m Click Run Analysis.

If you see a credit warning, you selected the wrong layer for your Analysis
Environments.

Note: Processing may take some time. The layer will appear in the Contents pane, but it will
be grayed out until processing is complete. Give this process some time before clicking the
web map.
The building footprints appear as a feature layer, represented as blue polygons in the map.

Copyright © 2022 Esri. All rights reserved. 6

 Imagery in Action

n After the layer finishes loading, save the map.

Step 3: Review extracted buildings

In this step, you will review the extracted buildings in your focus area. The output polygons
appear to capture many of the buildings in the area, but it is good practice to review the
results before continuing further analysis.

a In the Contents pane, for the AthensTestDeepLearning layer, click the Show Table button
to open the new layer's attribute table.
1. What two fields in the attribute table are important for your analysis?

__________________________________________________________________________________

__________________________________________________________________________________

b Click the Confidence field header and choose Statistics.

2. What minimum value is reported?

__________________________________________________________________________________

__________________________________________________________________________________

3. What average value is reported?

__________________________________________________________________________________

__________________________________________________________________________________

These values indicate that the confidence is very high in the polygons for this particular focus
area.

c Close the Statistics window and the attribute table.

d Zoom in to the lower-left area of the map and examine the quality of the extraction.

Copyright © 2022 Esri. All rights reserved. 7

 Imagery in Action

After reviewing the output, you can see that some buildings have not been extracted. After
considering the confidence values found in the feature layer, you see that none of the
buildings in the trees have been found. This omission is most likely due to the deep learning
tool not being confident in the buildings that it found in the trees and discarding them due to
the confidence threshold.

e Zoom to the full extent of the study area again.

Hint: Contents pane > Building Footprint Focus Area layer > More Options button > Zoom
To

Step 4: Re-extract building footprints

In this step, you will adjust the model arguments to find additional building footprints that
were not extracted in the previous process.

a Open the Detect Objects Using Deep Learning tool again.

Hint: Analysis > Raster Analysis > Deep Learning > Detect Objects Using Deep Learning

b In the Detect Objects Using Deep Learning tool pane, confirm that Choose Image Used
to Detect Objects is set to AthensTX_2018.

c For Choose Deep Learning Model Used To Detect Objects, choose Choose Deep
Learning Model.

d In the Choose Deep Learning Model window, click the My Content down arrow and

Copyright © 2022 Esri. All rights reserved. 8

 Imagery in Action

choose Living Atlas.

e In the results, select the Building Footprint Extraction - USA deep learning package.

f When the deep learning model arguments load, change the Threshold value from 0.9 to
0.7.
With the adjusted threshold, if the model is at least 70 percent certain that a feature is a
building, it will include it in the output.

g For Result Layer Name, type AthensFinalDeepLearning_<your initials>.

h Click Run Analysis.

Note: Recall that processing will take some time to complete, so do not click away from the
web map until the layer is loaded.

i In the Contents pane, drag the AthensTestDeepLearning layer above the

AthensFinalDeepLearning layer.

j For the AthensTestDeepLearning layer, click the Change Style button to open the
Change Style pane.

k For Select A Drawing Style, in the Location (Single Symbol) option, click Options.

l Click Symbols, and then change the Fill to green.

Copyright © 2022 Esri. All rights reserved. 9

 Imagery in Action

m Click OK to close the Symbols window.

n Click OK to close the Change Style window, and then click Done to save the style change.

o Zoom in to the middle of the focus area to view the results.

Copyright © 2022 Esri. All rights reserved. 10

 Imagery in Action

When zoomed in, you can see that the new blue polygons now represent many of the
buildings that were missing from the first analysis.

p Click one of the blue buildings to view the fields in the new feature layer.

Copyright © 2022 Esri. All rights reserved. 11

 Imagery in Action

In the preceding graphic, you can see that the confidence value for the chosen building is
84.44. This extracted building was previously not included, because the confidence value was
below 90. By adjusting the threshold value in the model argument, you allowed extracted
buildings with a lower confidence value to be included.
If this process was to be used for actual collection of extracted buildings, then the threshold
value might be further adjusted to find the optimal threshold.

q Save the map and close the browser.

In this exercise, you learned how to use prebuilt deep learning packages in ArcGIS Image for
ArcGIS Online to extract building footprints from an imagery layer. You also learned how to
improve results by modifying the model arguments.

Copyright © 2022 Esri. All rights reserved. 12

 Imagery in Action

Answers to Exercise Questions

1. What two fields in the attribute table are important for your analysis?
The Class field and the Confidence field

2. What minimum value is reported?

The minimum value is approximately 90.

3. What average value is reported?

The average value is 95.25.

Copyright © 2022 Esri. All rights reserved. 13