Deep learning applied to aerial imagery

Automatic change detection and feature extraction


News
July 10, 2018
Lee County property appraiser’s office turns to Pushpin change detection
April 24, 2018
Carson City accelerates its parcel change detection workflow using Pushpin’s deep learning platform
February 13, 2018
The Business Journal - Map-maker's last startup: Combat zones. This startup: Suburban homes
February 13, 2018
Maricopa County assessor leverages Pushpin’s deep learning platform to detect parcel changes and extract building footprints

Parcel Change Detection


95%
Accuracy
Pushpin provides next-generation property change detection leveraging a patent-pending, semi-automated workflow that consistently achieves human-level accuracy or better.
70%
Lower Cost
Pushpin flags changed parcels, indicates one or more change type, and divides parcels into geographic sectors. Pushpin delivers all of this at 70% lower cost than the competition.
4
Weeks
Pushpin’s accelerated workflow enables us to turnaround county change detection projects with hundreds of thousands of parcels in less than four weeks.

How it works


  • Pushpin developed the world’s largest parcel change detection training set. Each sample in the training set contains a parcel boundary with old and new high-resolution, leaf-off imagery. A map analyst trained in property assessment assigned a changed/unchanged label to each sample. For changed parcels, the analyst also assigned one or more change types (addition, construction, deck, demolished, flatwork, garage, new, other, pool, or shed).
  • Pushpin then used these samples to train, validate, and test a state-of-the-art convolutional neural network (CNN) to recognize changed parcels. Specifically, we use a custom network derived from the 50-layer residual network (ResNet50), which won the ImageNet Large Scale Visual Recognition Competition (ILSVRC) in 2015. Pushpin uses this custom neural network to automatically compare old and new images and classify each parcel as changed or unchanged.
  • Pushpin breaks large projects into batches of 500 parcels. Pushpin first uses the neural network to classify all batches. Map analysts then review 20% of the parcels in each batch. If the label agreement is 95% or greater, the batch is complete. Otherwise, a map analyst manually classifies the batch and another analyst reviews the result. Using this approach, we consistently achieve a project accuracy of 95% or better.
  • Competing approaches are completely manual and rely on building outlines. Unfortunately, relying on building outlines misses many changes that affect assessed value including non-building changes (e.g. a new deck, driveway, patio, pool, sauna, or spa) and changes that minimally impact the building outline (e.g. a teardown that utilizes the previous foundation).

Pricing


Pushpin charges 15¢ per parcel for our parcel change detection service. Sign up for our trial and we will analyze 1,000 parcels for free.

Impervious Surface Extraction


Catalog Properties
Pushpin leverages a patent-pending, semi-automated workflow that enables municipalities to rapidly catalog impervious surfaces for commercial and residential properties.
Calculate Utility Fees
Pushpin enables assessors to quickly and accurately calculate surface water management fees and storm water utility fees thereby increasing tax revenue.
Use Your Imagery
Pushpin can extract impervious surfaces information from the imagery that you already have. Pushpin supports 3-band (RGB) and 4-band (RGB + near infrared) imagery.

How it works


  • Pushpin developed the world’s largest impervious surface extraction training set. Each sample in the training set contains a parcel boundary with high-resolution, leaf-off imagery. A map analyst trained in property assessment outlined the impervious surfaces (e.g. buildings, decks, driveways, parking lots, patios, and sidewalks) within each parcel.
  • As with parcel change detection, Pushpin used these samples to train, validate, and test a custom neural network derived from ResNet50. Pushpin uses this neural network to automatically extract impervious surfaces from parcel imagery. Using this approach, we have achieved an average error rate of just 7% when calculating a parcel’s impervious surface area.
  • By comparing a parcel’s calculated impervious surface area to the one on the tax rolls, an assessor can quickly identify properties that may be underpaying for their surface water management fees or storm water utility fees.

Pricing


Pushpin charges 25¢ per residential parcel for our impervious surface extraction service. Sign up for our trial and we will analyze 1,000 parcels for free.

Case Studies


Hennepin County, Minnesota


434K
Parcels
13%
Changed
$29M
Tax Impact

Pushpin analyzed changes in 11,637 parcels in Hennepin county, Minnesota. The largest city in Hennepin county is Minneapolis. Pushpin used 1-foot, leaf-off imagery captured in the spring of 2009 and 2015 by Surdex Corporation for MetroGIS. Pushpin found that 13% of the parcels had changes affecting assessed value. Extrapolating to the county’s 434,208 parcels, this represents a potential annual property tax revenue increase of $28.9M. The following table summarizes the findings. For details, read the full report.

Change Type Frequency Tax Impact
Addition 3.1% $5.1M
Construction 0.4% $0.1M
Deck 0.8% $0.3M
Demolished 0.1% $0.0M
Flatwork 8.1% $1.4M
Garage 0.4% $0.5M
New 2.7% $21.2M
Other 0.4% $0.1M
Pool 0.3% $0.1M
Shed 0.3% $0.1M

Wake County, North Carolina


358K
Parcels
9%
Changed
$15M
Tax Impact

Pushpin analyzed changes in 1,000 parcels in Wake county, North Carolina. The largest city in Wake county is Raleigh. Pushpin used 6-inch, leaf-off imagery captured in the spring of 2010 and 2013 by Sanborn Map Company for NC OneMap. Pushpin found that 9% of the parcels had changes affecting assessed value. Extrapolating to the county’s 358,465 parcels, this represents a potential annual property tax revenue increase of $15.3M. The following table summarizes the findings. For details, read the full report.

Change Type Frequency Tax Impact
Addition 2.0% $2.7M
Construction 0.3% $0.1M
Deck 0.4% $0.1M
Demolished 0.6% $0.0M
Flatwork 4.8% $0.7M
New 1.8% $11.6M
Pool 0.2% $0.1M
Shed 0.7% $0.1M

Milwaukee County, Wisconsin


162K
Parcels
5%
Changed
$1M
Tax Impact

Pushpin analyzed changes in 1,000 parcels in Milwaukee county, Wisconsin. Pushpin used 6-inch, leaf-off imagery captured in the spring of 2013 and 2015 by Pictometry. Pushpin found that 5% of the parcels had changes affecting assessed value. Extrapolating to the county’s 161,536 parcels, this represents a potential annual property tax revenue increase of $1.4M. The following table summarizes the findings. For details, read the full report.

Change Type Frequency Tax Impact
Addition 0.3% $184K
Construction 0.2% $19K
Demolished 0.1% $0K
Flatwork 3.3% $213K
Garage 0.2% $103K
New 0.3% $872K
Other 0.3% $10K
Shed 0.4% $13K

Hartford County, Connecticut


347K
Parcels
8%
Changed
$3M
Tax Impact

Pushpin analyzed changes in 1,000 parcels in Hartford county, Connecticut. Pushpin used 1-foot, leaf-off imagery captured by Photo Science in the spring of 2012 and 3-inch, leaf-off imagery captured by Sanborn in the spring of 2017. Pushpin found that 8% of the parcels had changes affecting assessed value. Extrapolating to the county’s 347,225 parcels, this represents a potential annual property tax revenue increase of $3.1M. The following table summarizes the findings. For details, read the full report.

Change Type Frequency Tax Impact
Addition 0.7% $924K
Deck 0.4% $111K
Flatwork 5.7% $792K
Garage 0.5% $556K
New 0.1% $625K
Other 0.1% $7K
Pool 0.2% $83K
Shed 0.6% $42K

Brazos County, Texas


65K
Parcels
14%
Changed
$5M
Tax Impact

Pushpin analyzed changes in 1,000 parcels in Brazos county, Texas. The largest city in Brazos county is College Station. Pushpin used 6-inch, leaf-off imagery captured in the spring of 2013 and 2017 for the Texas Natural Resources Information System (TNRIS) Strategic Mapping Program (StratMap). Pushpin found that 14% of the parcels had changes affecting assessed value. Extrapolating to the county’s 65,385 parcels, this represents a potential annual property tax revenue increase of $5.2M. The following table summarizes the findings. For details, read the full report.

Change Type Frequency Tax Impact
Addition 2.2% $547K
Deck 0.1% $5K
Demolished 2.3% $0K
Flatwork 2.4% $63K
Garage 0.7% $146K
New 3.7% $4,355K
Other 0.5% $7K
Shed 3.2% $42K

About


Company


Founded in 2015 and based in Minneapolis, Pushpin believes that people and computers intelligently combined can solve difficult mapping challenges better, faster, and cheaper than the alternatives. Pushpin works with tech-forward customers and partners to dramatically increase automation, accelerate workflows, and decrease costs. Pushpin applies patent-pending deep learning algorithms to aerial and satellite imagery to identify parcel changes, extract building footprints, estimate impervious areas, and more. Pushpin’s customers include Carson City, Nevada; Hillsborough County, Florida; Lee County, Florida; Maricopa County, Arizona; and Pope County, Minnesota.

Pushpin’s founder and president is Randy Milbert. Randy is an MIT graduate who previously served as founder and president at Primordial, which developed off-road route planning software for the military. In 2013, Randy sold Primordial to Polaris Industries (NYSE:PII).

Pushpin has world-class computer science and machine learning expertise with an emphasis on cutting-edge deep learning. Two of the company’s principals—Randy Milbert and Andreas Robinson—each have 15+ years of experience with advanced algorithm development and winning competitive research and development contracts from the United States Army, Air Force, and DARPA. Pushpin’s patent-pending approaches for parcel change detection and impervious surface extraction are more accurate, less expensive, and an order of magnitude faster than existing solutions.

Team


The following are Pushpin’s principals. Pushpin also has a team of map specialists who assist with training set creation as well as support staff including business development consultants and graphic designers.

Randy Milbert

Founder and President

Pushpin's founder and president is Randy Milbert. Randy graduated from the Massachusetts Institute of Technology (MIT) in 2000 and founded Primordial in 2002. At Primordial, Randy led the development of the company's Ground Guidance® off-road navigation software and Ooze® crowdsourcing geographic information system (GIS) platform. He secured more than $15 million in contracts from the United States Army, United States Air Force, Missile Defense Agency, Rockwell Collins, General Dynamics, Lockheed Martin, Polaris Industries, and others. He received six issued and three pending patents. In 2012, the Army acquired a government-wide license for Ground Guidance. In 2013, Polaris Industries (NYSE:PII) acquired Primordial. In 2015, Randy left Polaris Industries to found Pushpin.

Andreas Robinson

Chief Technology Officer

Pushpin's chief technology officer is Andreas Robinson. Prior to joining Pushpin, Andreas was an early hire and senior engineer at Primordial where he led multimillion dollar Army, Air Force, and DARPA projects including developing an aerial vehicle-based geospatial tracking system, an augmented reality platform for dismounted soldiers, a learning-based navigation system, and a machine learning application for extracting features from satellite imagery. Prior to Primordial, he developed neural networks for flush-air sensor systems at Goodrich Aerospace (now United Technologies). He has a masters degree in computer science from the University of Minnesota from 2004 and a physics and math degree from Reed College from 1999.

Grady Hannah

Vice President Business Development

Pushpin's vice president of business development is Grady Hannah. Grady is an 18-year technology sales and business development veteran, spending 15 years in senior roles in tech-forward Silicon Valley companies. These markets include manufacturing industries, real-time entertainment software, Linux and open source software, and alternative energy. Grady's experience in the assessor market comes from a sales role for ground-based image capture company Facet Technologies. Grady's role at Pushpin is to sell and market Pushpin's image-based change detection and feature extraction capabilities to all relevant markets including county, city and state governments, emergency services, GIS, ISV partners, and discrete map creation applications.

Scott Sherwood

Consultant

Scott Sherwood retired from EagleView (formerly Pictometry) in 2017. He joined Pictometry as a startup in 2000 and served in various sales and marketing positions over 17 years including vice president of sales (northern region), vice president of corporate marketing initiatives, and vice president of sales enablement leading up to the merger with EagleView in 2013. While at EagleView, he was responsible for product development strategy for the public safety market. Prior to Pictometry he was with the Center for Governmental Research Inc. in Rochester, NY for 26 years where he served as the director of technical services responsible for project management in environmental planning, land use planning, and public safety applications of GIS and GPS.

Customers


Testimonials


Pushpin is applying to real property assessment the same kind of artificial intelligence algorithms that Google uses in its autonomous cars. Our testbed was a set of four areas of the county, each of which presents a different challenge for change detection via aerial photography. After we examined these small pilot areas with this constantly learning technology, we believe our office has a new opportunity to improve the accuracy, turnaround time, and cost for detecting a variety of changes.

Eric Bails

Chief Technology Officer, Maricopa County Assessor, Arizona
Pushpin delivered high-quality results in days versus months and did so at one-fifth the cost of the competition. Pushpin is applying to property assessment the same kind of artificial intelligence algorithms that Google used to beat world champion Go players. The result is a dramatic improvement in accuracy, turnaround, and cost—and a phenomenal return on investment.

Dave Dawley

Assessor, Carson City, Nevada
We fly new aerial photographs of Lee County every year to help with our parcel inspection and valuation process. Pushpin allows us to leverage that investment by identifying changes in parcels without costly on-site inspections. Having a full accounting of new construction, additions and demolitions is important in developing a fair and equitable tax roll for everyone. Pushpin is a great new tool we plan on utilizing to increase the accuracy of our data collection process.

Ken Wilkinson

Property Appraiser, Lee County, Florida

Contact


Pushpin is headquartered in Minneapolis, Minnesota.

Feel free to contact us anytime at 651-216-6941 or randy.milbert@pushpin.us.

Try for Free


Pushpin will analyze 1,000 parcels in your county for free. Please use the following form to provide your contact information. We will follow up to obtain the relevant parcel boundaries and imagery. Once we have that, we will deliver final results within five business days.