Setting Stormwater Rates Using Machine Learning

City of Mineral Wells, Texas, Uses Image Classification in Stormwater Rate Study

stormwater drain

By Tak Makino


The City of Mineral Wells, Texas, is a community of approximately 15,000 people west of the Dallas-Fort Worth metroplex. The city has experienced numerous floods in recent years, including significant damage in the 2015 Memorial Day floods. Its stormwater infrastructure needs repair, maintenance and expansion that is currently not in the city’s budget. To address this shortfall, the city contracted with Lockwood, Andrews & Newnam, Inc. (LAN), a national planning, engineering and program management firm, and NewGen Solutions, a management consulting firm, to perform a stormwater utility rate study.

Stormwater utility fees are a popular method to build capital for constructing, maintaining and expanding a city’s stormwater infrastructure. A stormwater utility fee provides a transparent, usage-based method of fee collection. Much like how a municipal utility bills for drinking water on a consumption basis, a stormwater utility fee is also billed on a consumption basis. Under a stormwater utility fee, individual or commercial properties that send more water to the stormwater system are responsible for a greater share of the utility cost.

Impervious surface coverage provides a proxy for runoff contributed to the stormwater system. In other words, the more impervious surface or pavement on a lot, the greater the fee is to send that runoff water to the stormwater drainage system.

Don’t Have the Data?

One of the major components that goes into calculating a stormwater utility fee is an impervious surface database. An impervious surface database tracks how much impervious surface exists in the form of buildings or pavement on each plot of land in a community. Currently, the city of Mineral Wells does not have an impervious surface database. As purchasing a ready-made impervious surface database was cost prohibitive, LAN generated an impervious surface dataset for this study by using novel machine learning techniques.

Generate the Data

Companies like Google and Facebook use machine learning to help improve user experiences. Supervised machine learning is a computational technique in which an operator provides the computer a set of training examples for a particular class of output values. For example, Facebook recognizes your friend John Doe because many people have uploaded pictures of John (the training set) and then tagged John in the photos (the output value). Facebook now recognizes John in new photographs because machine learning algorithms have learned the characteristics that define John’s face.

Rather than using machine learning to identify faces, LAN applied machine learning to identify areas of pervious and impervious land cover. For example, rather than telling the computer that an area is covered in grass, examples of other grassy areas are provided to the extent that the computer learns what grass looks like and is then capable of identifying grassy areas it has never seen before.

The project team identified different machine learning algorithms

The project team identified different machine learning algorithms that could be compared against each other to produce a high-quality impervious surface dataset.

Developing the Data Set

Using machine learning to classify aerial imagery is a well-established practice in the remote sensing industry today. Such machine learning tools are standard functions in commercial GIS software. LAN developed in-house methods that leveraged the strengths of the machine learning algorithms to quickly and accurately develop an impervious surface dataset that fit the client’s needs.

Rather than relying on a single binary (impervious vs. pervious) training dataset, the project team identified different machine learning algorithms that could be compared against each other to produce a high-quality impervious surface dataset. Machine learning algorithms are not one-size-fits-all. LAN’s remote sensing experts identified methods that took advantage of the strengths presented by different machine learning algorithms to produce an accurate product.

A police lineup provides a reasonable analogy for discussing machine learning algorithms and aerial imagery classification. In the case of a police lineup, a criminal suspect and several other individuals of similar appearance are shown to an eyewitness. The addition of similar-looking people ensures that the eyewitness’ identification of the suspect is sufficiently challenged and can be presented as evidence in court. Similarly, in the case of machine learning, additional training classes also make class identification more difficult. However, unlike the case of a police lineup, the machine learning operators are aware of the identity of the suspect before running the analysis (the operators already know what grass looks like). The question becomes how to best train the machine learning algorithm to identify land cover classes. In the case of a police lineup, this would be analogous to identifying methods to train an eyewitness to identify a suspect.

For the stormwater study, the project team created several datasets to train the machine learning algorithms to discriminate between land cover classes that were similar in appearance. For example, aerial imagery of dirt and pavement can appear similar – both can exhibit shades of grey or brown. Critically, for the purposes of this study, dirt is a pervious land cover that absorbs water and pavement is an impervious land cover that leads to runoff. By first optimizing identification between difficult-to-discriminate land cover classes before introducing other land cover classes, the additional classes are then less “confusing” to the algorithms. Returning to the police lineup analogy, if the police first coach an eyewitness to identify a suspect in isolation, without the confusing influence of additional individuals, the eyewitness will be better equipped to identify that suspect in a fully populated police lineup.

After using machine learning to generate the impervious surface dataset, LAN determined that the typical single-family residence in the city has approximately 2,600 sq-ft of impervious surface. This value of 2,600 sq-ft is referred to as the Equivalent Residential Unit (ERU) and forms the basis of the stormwater utility fee.

The Numbers

The City of Mineral Wells currently bills all utility customers $2.50 per ERU for the stormwater utility, regardless of property size or impervious surface coverage. This is at the lower end of the $2.50 – $6.50 per ERU range seen in similar communities in Texas. NewGen Solutions took the machine learning dataset provided by LAN, performed financial projections and concluded that to fully fund its stormwater infrastructure for the next five years, the city requires a $3.59 fee per ERU. This rate will cover all anticipated repair, maintenance, expansion and other costs associated with providing quality stormwater infrastructure to the residents of Mineral Wells.

The utility fee has yet to be adopted by the city council. Under one proposed billing scheme, all single-family residential utility customers will pay a flat fee of $3.59 per month to the stormwater utility. All commercial and other non-single-family-residential utility customers will pay on an ERU basis. It is possible that a different billing scheme will be adopted by the city council, but the proposed scheme is similar to utility fee practices observed in similar communities.

This proposed change from $2.50 to $3.59 will increase the rate for all single-family residential utility customers and most commercial customers. To allow residents to plan and budget for the anticipated rate increases, it is possible that the city may announce a phased implementation schedule, starting by switching to ERU-based billing at a reduced rate. Over time, the rate would increase to the $3.59/ERU rate that fully funds the utility. The phased increase implementation schedule allows for planning and budgeting. However, it is possible that the city council may adopt a different implementation scheme.

Where Are We Headed?

Machine learning is a nascent field that shows tremendous promise. The research and technology sectors have embraced machine learning. Other industries have been slower to adopt machine learning. The applications of machine learning are myriad. Image classification is just one of the many ways machine learning can be used in water finance applications. As the adoption of machine learning becomes more widespread, additional applications will be found for this technology, improving efficiency, accuracy and helping to further leverage existing tools and methods.


Tak Makino is a flood mitigation manager at Lockwood, Andrews & Newnam, Inc. (LAN), a national planning, engineering and program management firm. Makino specializes in floodplain management, land use planning, spatial data analysis, data management and natural hazard mitigation. He can be reached at tmmakino@lan-inc.com.

Leave a Reply

Your email address will not be published.

*