Satellites Guide Habitat-Friendly Solar
Have you ever taken a road trip across Long Island? Parked your car in one of numerous parking lots covering the island? Have you ever wondered just how many parking lots there are? If you’re like me, probably not. It had never occurred to me to locate every single parking lot on Long Island, but, as an intern with Defenders of Wildlife’s Center for Conservation Innovation, that is exactly what I attempted to do this summer! Why? Because some of them may be good places to add solar panels.
One of nine states in the Regional Greenhouse Gas Initiative, New York has some of the nation’s more aggressive renewable energy goals, including a target to generate 50 percent of the state’s electricity through renewable sources by 2030. Long Island plays a key role in that effort, boasting the state’s first offshore wind farm and New York’s three largest solar farms. Recently, the Long Island Power Authority (LIPA) — which supplies electricity to residents of Long Island — committed to bringing an additional 800 megawatts of clean energy, enough to power over 300,000 homes, to the grid by 2030. Solar energy can play a critical role in fulfilling LIPA’s commitment due to Long Island’s high solar potential.
Defenders of Wildlife has partnered with The Nature Conservancy to develop a “solar roadmap” for Long Island to prioritize areas for solar energy development. The question is, where do you put all those solar panels? Ideally, we want to find places that are near existing infrastructure (like transmission lines) but do not require the conversion of large swaths of open space, which can destroy wildlife habitats and recreation opportunities. That’s where parking lots come in. Parking lots are often near population centers, and solar panels built as carports over parking lots ensures that the primary purpose of the land (parking) is not compromised. They can even provide extra benefits to drivers in the form of protection from sun and rain. Critical to Defenders’ mission, building solar panels over parking lots involves little to no disturbance of natural wildlife habitat.
One of the parts of the solar roadmap is to assess just how much solar energy could be put on parking lots, and at what cost. But first, we had to find the parking lots, preferably without individually tracing them all on a map. My challenge was to locate and map all the parking lots on Long Island so that these sites could be systematically prioritized. This was a daunting task made more manageable by the availability of satellite imagery, machine learning tools, and the Google Earth Engine platform. I used a technique called supervised classification to train a machine learning model to recognize the characteristics of different landcover types (e.g. pavement, building, forest, agriculture) in the pixels of a satellite image, and then find those same characteristics in new pixels to predict their landcover types.
Although there are many different methods of image classification, this project focused on two machine learning techniques: random forest classification and deep neural networks (DNN). A random forest classifier is like a standard decision tree in which data is used to identify thresholds that separate observations (pixels in this case) into different classes. DNNs are one of the many new artificial intelligence approaches designed to mirror the human brain by creating a series of layered algorithms that recognize patterns in data.
Random forest and DNN classifiers require lots of data to learn characteristics of pixels in both the categories of interest, and other unwanted classes. I collected this model training data using the known location of 3,808 parking lots and 5,418 buildings from across Suffolk and Nassau counties, along with 6,029 randomly generated points in areas of forest and grass landcovers. These points were used to sample radar data images collected by satellites in four different seasons. The radar data provides information about the physical structure of objects to help distinguish between rooftops and parking lots, which look similar.
The random forest classifier identified paved surfaces (including parking lots), buildings, forests, and grass with an overall accuracy of 90.7%! Including the radar data improved the classifier’s ability to differentiate between pavement and buildings, but the DNN model performed better without it. Including seasonal variation seemed to have minimal impact. On a pixel-by-pixel basis, parking lots look identical to roads in satellite images, so I eliminated all areas identified as paved that overlapped with roads. This left me with all the parking lots on Long Island!
Beyond finding all these parking lots, I wanted to compare the efficacy of our different models and determine what information is necessary to be successful. Determining the best model and set of input data not only allowed us to most effectively identify all the parking lots on Long Island, it also provided a sound foundation for future analyses of potential sites. For example, should it be determined that, in addition to parking lots, open areas surrounding airport tarmacs might be suitable sites for solar installation, our established model structure could be easily tweaked to identify those features. Or it could be used to detect strip mines in Pennsylvania. Or logging sites in Oregon. Tools such as these will allow Defenders’ Center for Conservation Innovation to be better informed about the spatial spread of threats facing species across the country. They offer a way to prioritize conservation and restoration efforts, as well as minimize the impact of the built environment on fragile habitat.
More immediately, however, this project will be used in conjunction with opinion research of Long Island residents to locate optimal sites for solar panel installation, a crucial step towards achieving New York’s renewable energy goals.