Identifying Wetlands Using AI and Deep Learning

This article was written by EPRI staff

Wetlands are an important asset and ecosystem within our landscapes. In addition to their natural flood protection, wetlands are home to more than one-third of the United States’ threatened and endangered species. To ensure their protection, utilities must be able to accurately identify the location and characteristics of wetlands and watersheds in their regions of operation. This information, which is shared with regulators who permit impacts to wetlands and streams in the United States under Section 404 of the Clean Water Act, is used during the planning and siting phases of renewable energy projects, transmission lines, or other infrastructure projects.

To help with this need, researchers in the Ecosystem Risk and Resiliency program at the Electric Power Research Institute (EPRI) began pursuing a new effort to improve and refine wetland identification. The effort aimed to improve wetland identification accuracy and also improve efficiency.  

Wetland identification is often a resource-intensive effort. During the planning stage, the effort can be fraught with errors, as national-level datasets like the National Wetlands Inventory (NWI) and National Hydrography Dataset (NHD) have limitations including omissions of wetlands sites and mapping of wetlands that no longer exist. In many regions, data is also out of date, with maps and information from the 1980s. Understanding these challenges, EPRI collaborated with the Chesapeake Conservatory to explore a new process which could leverage advances in artificial intelligence (AI) and deep learning models, satellite photo imagery, and light detection and ranging (LIDAR) remote sensing to improve the accuracy and quality of wetland information. If successful, the process could dramatically reduce the time and resources needed for initial assessments, while increasing confidence in the precision of these identifications. 

EPRI worked closely with utility members across the U.S. to determine the initial functionality needed for this model. These discussions led to a focus on delivering a highly accurate inventory of wetlands and streams through an automated process. Future iterations could provide insight and understanding into the characteristics of identified wetland sites. 

To build a more accurate inventory, the Chesapeake Conservatory’s Conservation Innovation Center (CIC) proposed an AI architecture that uses a deep learning model (called U-Net) and data inputs that allow for scalability in wetland identification. The approach allowed the model to take advantage of the power of deep learning and incorporate object-based contextual information which can reduce the need for expensive or heavily engineered data inputs.  “The method uses high-resolution, free, and publicly available data inputs with an open-source U-Net design to make this approach scalable,” explained Susan Minnemeyer, vice president of Technology at Chesapeake Conservancy.

The deep learning model uses satellite imagery that captures both visible light and light bands beyond the visible spectrum, and three-dimensional LIDAR data that indicates elevation and inundated soils. Initial results showed the model worked particularly well in areas that are often challenging to map accurately with photoimaging alone. For example, areas where forest canopies or other overhead covering conceal wetlands or streams were accurately identified by the AI model. 

While the model was not intended to replace field visits (which are used to confirm wetland delineation in permits), pilot phases showed significant value by drastically improving the efficiency and accuracy of initial assessments as compared to today’s process. “As COVID-19 restrictions reduced field visits, the importance and application of this tool became even more apparent,” said Becca Madsen, senior technical leader at EPRI.

When it came to utility partners, the model’s success was evaluated by several key metrics. The first, and perhaps most obvious, was accuracy of prediction. This includes both accurate identification of areas that are wetlands, and those that are not. Overall accuracy was then evaluated to determine the percentage of the model’s false identifications.  

“We had enormous interest in this project from the start,” explained Vin Pezzullo, manager of Environmental Studies and Remediation at the New York Power Authority (NYPA). “NYPA is working towards becoming the first digital end-to-end utility and this model fit right in with that vision.” 

From the outset, NYPA knew that use and adoption of the model would only be possible if it was able to meet its high-performance standards, including at least 90 percent accuracy In a demonstration in fall of 2021, the project achieved NYPA’s goal by achieving 94 percent accuracy.

“We were really thrilled with this initial performance and outputs,” said Madsen. “It validated that this approach really has potential to transform the current wetland assessment process.” 

NYPA plans to begin testing the model by incorporating it into their existing wetland screening process this year. “If we can show this approach provides a very accurate assessment quickly, and one that we can trust and confidently replicate, I ’m hopeful we’ll gain the support for its adoption in the regulatory arena,” explained Pezzullo. 

Madsen, EPRI, and CIC colleagues hope this year provides new opportunities for collaboration and application of the model at scale. “We see the potential for enormous efficiencies by taking this model and applying it to regions or multi-state watersheds like the Chesapeake Bay.” 

In the meantime, the project team shared outputs for the first three geographic areas - Delaware, Minnesota, and New York – and a video tutorial of how to use the application and outputs with the public.  

EPRI’s Nalini Rao, who leads the Ecosystem Risk and Resilience program, shared her optimism about this model and the opportunity to expand its application. “The pilot results have been very promising and we’re optimistic about the potential for this model to significantly improve the wetland assessment process across the United States and, ultimately, protect and preserve these important landscapes.” 

Learn more about EPRI and Chesapeake Conservatory’s deep learning model by visiting this interactive webpage or contacting bmadsen@epri.com