Leveraging Automation to Scale Urban Geospatial Production
Radiant Solutions utilizes the latest automation techniques and data products with satellite imagery to create powerful geospatial products and services that can be applied anywhere in the world. Leveraging our decades of expertise of working with our government and commercial customers, we are creating mapping tools to enable collaborative mapping and scaled geospatial production workflows across a distributed user base. Our production teams are focusing on feature extraction in urban areas — the places where more than half of the world’s population lives and where 70% will reside by 2050. The production of dense urban feature data is critical for vital services and time sensitive missions in these locations for humanitarian assistance, disaster response, and sustainability planning. The density of road, building, and POI information in urban areas makes urban feature data production labor intensive and challenging to scale over large areas.
Radiant Solutions is scaling dense urban feature production in three ways: (1) leveraging automation to produce urban feature geometry from satellite imagery; (2) developing and leveraging collaborative feature extraction tools that enable volunteer mappers to efficiently extract, enrich, and ultimately deliver high quality urban map features; and (3) scaling mapping crowds to enable to enable a proper urban feature data production/finishing capacity. In this first blog of a continuing series, the Foundation GEOINT team will show how Radiant Solutions leverages automated data products to drive production efficiencies for dense urban feature dataset production.
History has demonstrated that investments in automation can drive production costs down and increase production capacity. Most notably is Henry Ford, who pioneered this tactic to bring mass production of the Model-T to marketplace. History has also shown cases where automaton doesn’t always achieve these desired results. Consider General Motors, who ineffectively spent billions in the mid-1980s to automate assembly tasks that humans simply did better. Similarly, Elon Musk fell into a trap of over-automation in the Spring of 2018, stating that “Excessive automation at Tesla was a mistake”, citing cases where human production was undervalued in the Model 3 assembly line design. Automation has proven that it can reduce production costs and increase production capacity, not only for producing cars but for mapping data as well. That said, sometimes the state of the art in science and engineering are not enough to provide a viable automated solution, and therefore humans-in-the-loop are needed.
When it comes to automation for urban feature production, we have two simple rules:
- Automation must be more efficient than manual methods.When data tools or techniques do not conform to this rule, we rely on conventional feature extraction and enrichment methods and/or hybrid solutions.
- Automated products must be validated and reviewed before, during, and after integration into the database.It is important to ensure automated products don’t include collection anomalies (i.e. poorly formed geometry, missing geometry) prior to ingest, and, upon integration, database errors are not introduced (i.e. duplicate or conflicting geometry). Facebook describes a validation/review workflowas one example on how automated data products are integrated into OpenStreetMap (OSM).
The building footprint is more than just a polygon feature — it serves as the base geometry for key points of interest, navigation way points, and physical street addresses (i.e. 1 Main Street). Conventional feature extraction specifications typically limit building footprint collection to key points of interest with rules to enforce minimum distance/maximum density to limit the number of building footprints to a manageable number. Generally speaking, full extraction of building footprints requires significant person-hours to complete and is typically reserved for small urban areas and villages. Complete building extraction specifications assume there is an available feature extraction team that can scale to meet production timelines. For larger cities, a building footprint extraction sample rule (minimum distance, specific building types) is often necessary to keep manual feature extraction cost at an affordable level.
The Radiant Solution dense urban mapping specification requires that building footprints must:
- Be extracted at a completeness rate of 95%
- Be extracted orthogonally (squared corners when appropriate)
- Align to the base (foot) of the building
- Have a “building” feature tag
- Model adjacent buildings (i.e. row houses, townhouses)
- Buildings must be tagged with the date of imagery
- Buildings must be extracted from an image source of less than one year
Let’s suppose an urban area with 279,000 buildings needs to be mapped. With an average extraction rate of 120 buildings/person-hour (Ref: Table 1), it would take an estimated 14.5 person-months to extract the buildings in this urban area. Assuming analysts maintain the average production rate while with an entry level salary, the total labor cost would be approximately $35,000. This assumes the production team could maintain a consistent extraction rate producing quality footprints for the length of the campaign — something we have found not to be realistic at this scale. In the end, a $35,000 cost may be too expensive (skilled labor rates too high), or the production timelines too long (if there is not enough skilled labor) to complete the project in a timely manner.
In the absence of automation, Radiant Solutions would have to dramatically scale back the production specification to sampled buildings to keep cost down and production timelines to a reasonable level. Keep in mind that a down-sampled urban feature extraction specification limits the value of urban feature datasets for routing and other location-based services given the limited feature density. Given the unfavorable options of expensive manually extracted urban data or the cheaper option of a sampled dataset, Radiant Solutions has chosen to go a different route and turn to automation to scale up our dense urban feature production.
Location based maps and analytic services (such as search, geocoding, and routing) require high fidelity building footprint information. Radiant Solutions, ala Henry Ford, has turned to automation in order to acquire complete building footprint geometry for our production. We leverage Ecopia Building Footprints Powered by DigitalGlobe, a dataset produced by our partner Ecopia.ai through machine learning enabled automation capabilities applied to DigitalGlobe high-resolution satellite imagery. This building footprint product has the following characteristics:
The Ecopia.ai production capacity currently scales up to produce 40 million curated building footprints per month based on the latest high-resolution DigitalGlobe satellite imagery. This system was leveraged to map all building footprints in Australia and has recently applied to map all building footprints in the United States. This data product yields unmatched quality and production/scale characteristics through manual “curation” of building footprint training datasets, QA/QC of the automated results, and where necessary manual extraction. Recently the US dataset was leveraged by our GeoHive team to summarize Hurricane Michael storm damage (with 5x cross validation) for over 7,100 buildings in less than three hours.
Production cost for the Ecopia building footprints is less than the estimated cost of the manual notional example above for 279,000 features. From a timeliness standpoint, we have seen production delivery timelines of this size or greater to take days to weeks (an order of magnitude faster) for curated building footprint datasets. For obvious economic reasons, we no longer employ personnel to manually extract 2D building footprints in bulk — not even interns! We do license an increased amount of product from Ecopia & DigitalGlobe to underpin our enriched urban feature data products.
At Radiant Solutions, we leverage automation to achieve production efficiencies, lower cost, and maintain high levels of quality. In the upcoming blog posts, we will highlight how we are developing, integrating, and extending collaborative feature extraction tools and creating high quality geospatial data products leveraging modern tradecraft.
Contact us if you are interested in our Urban Feature Extraction services. Also we are also always on the lookout for talented developers, engineers, and analysts. If you are interested in a new career opportunity, we have immediate openings for Software Development Engineers, Senior Software Engineers, Systems Engineers, and DevOps Engineers.
Also, if you’re looking for an internship and don’t want to manually extract building footprints in mass, contact us.
Originally published at medium.com on November 6, 2018.
