The rise of Artificial Intelligence (AI) is changing the way our society looks and interacts with data. AI, combined with the expanding tools of Geographic Information Systems (GIS), allows for dramatically more efficient data visualization and cost-effective analysis.
As part of my Masters of Urban Planning Capstone project at San Jose State University, I was tasked with identifying the acreage of surface parking lots in our North San Jose study area. This meant finding parking lots on nearly 4,000 acres (over six square miles) of land or identifying hundreds of pools from the view of an airplane window.
Rather than use Google Earth or spend many hours exploring the site area, I made a map using ArcGIS Deep Learning Model, specifically the Parking Lot Classification Tool. This analyzes a raster image file (a digital image file, like JPEGs or TIFFs), pixel by pixel, creating a separate layer designating areas as parking lots as appropriate.
In simple terms, the model looks for color patterns in the picture and provides a data output for visualization, similar to tasking a person with sorting a bag of colored beads. All the grey (parking lot) goes in one pile, and all the others in another pile. It took me just a few hours of developing and modifying a task, then I let the computer run for a few days, to determine that approximately 16% (640 acres, 1 square mile) of the project area is covered by surface parking.
Parking Lot Identification
The mapping process, much like a GIS map itself, is made of layers. Before starting in ArcGIS Pro, I conducted background research. There are many different models, some Deep Learning models are trained to detect objects like cars or to detect pixels for classification, while others detect change over time.
To begin, I used a TIFF file, or a high-quality raster image file, provided by the City of San Jose. In ArcGIS Pro, I imported the raster file and the study area boundary layers, matching the image to the shapefile. After this was the most difficult part: training the model for more accurate results. In order for the program to recognize the various grey patterns and shades of the parking lots – as opposed to the grey of streets or rooftops - I developed training samples for the Parking Lot Classification model. This required creating a new feature class and manually outlining small parking lot section polygons over the aerial image raster file layer.
Figure 1: Parking Lot Classification Model Training Sample Polygons
Once exportation of the training data was complete, I used the Classify Pixels Using Deep Learning tool. After about 70 hours of background runtime using a gaming PC, the preliminary map was complete, more accurately identifying parking lots in purple.
To clean up Figure 2 to Figure 3, I used the Dissolve and Smooth Polygon tools to combine the yellow and purple layers and smooth out the jagged edges of the polygons to create the final parking lot map below.
Figure 4: Map of North San Jose with parking lots shown in a yellow color (Map produced by Hannah Meeks, Fall 2023 URBP 295, City of San Jose Open GIS Data Portal: Santa Clara County 2022 Imagery basemap. Valley Water Open Data Portal: Rivers for Study Area Boundary. Self-Created: Highway Shields.
Applying AI to Planning Practice
The study area is primarily made up of industrial/commercial facilities, home to many large tech companies such as Samsung, Cisco, and PayPal. There is a small amount of mixed-use commercial and residential development as well.
The parking lot analysis showed that a tremendous amount of the study area was dedicated to vehicle parking: land that is essentially storage, and un-activated outside of rush hour or folks grabbing lunch.
As transit options continue to improve, and the private car (or pickup truck, or SUV) loses its primacy in the Bay Area, parking lots represent development potential. The transformation of existing parking lots provides fantastic infill opportunities. This could mean civic or public space – including desirable green spaces – as imagined by Gensler’s architecture firm, Los Angeles cultural center The Mod, as seen in this Curbed article Parking Garages Are Getting a Second Life as Places for People.
Figure 5: Modular buildings inserted into existing parking garage. Image source Curbed
Figure 6: Parking garage used as public space. Image source Curbed
It could also mean much more affordable housing. Not just specially-built affordable housing, but simply dramatically reducing the cost of market-rate housing. According to NPR, parking spaces “can add major costs to building new housing: a single space in a parking structure can cost $50,000 or more.”
Finally, parking spaces too often contribute to urban heat islands. As described in my colleague Laylonni Laster’s article, “heat islands are areas where structures such as buildings, roads, and other infrastructure that absorb and re-emit solar heat are highly concentrated and natural landscapes are minimal…Materials, such as pavements or roofing, tend to absorb and emit solar energy rather than reflect it…Therefore, these materials can contribute to the increase in local temperatures and foster heat island development.”
Urban heat islands pose serious health risks, including heat exhaustion and heat-related illnesses for vulnerable communities, including seniors, and elevates energy consumption, which increases air pollutants and greenhouse gas emissions. The pavement decreases water quality as polluted runoff collects in streams and rivers and impacts aquatic life.
Each of my telltale grey pixels was an area that contributed to urban heat and the health risks they incur. Identifying – and eventually removing - these parking lots can help promote climate change initiatives by policymakers with the addition of plants, trees, and other green spaces to reduce climate change effects.
How does GIS answer planning questions?
When I first began this GIS Deep Learning project it felt overwhelming, but as I researched and accomplished each step, it was attainable. Our project would never have discovered the impact of surface parking lots in this area without the Deep Learning tool. For GIS technicians, it is important to understand when we must execute the GIS steps ourselves versus when we need additional GIS tools in order to solve the community’s needs.
AI is not a solution to all urban planning problems, not by a long shot. But what careful application of computer learning can do is increase the efficiency of some tasks, and deliver more careful analysis to a community’s elected leaders. By taking advantage of built-in processes like the Deep Learning tool, M-Group can work with jurisdictions to provide efficient GIS solutions to communities to ask questions they might not have believed they could afford to ask.
By Hannah Meeks, Assistant Planner