Remote Sensing Series: GIS Layers for State and Local Government

Series Part 1: Impervious Surface Mapping

Introduction

Impervious surfaces, or more explicitly, man-made surfaces such as roads, sidewalks, buildings, and parking lots, make up more and more of the urban landscape as communities continue to experience growth and make necessary infrastructure improvements. The environmental impacts that result from an increase in impervious surfaces range from water quality reduction, increased flood potential, wildlife habitat disruption, and many other challenges for communities to address.

The first step in addressing these challenges is to qualify and understand the current extent of impervious surface coverage, and also to quantify coverage over a particular area of interest (AOI). In addition, municipalities tasked with such projects must also do so within various time and budget constraints.

Low-cost or sometimes freely available data coupled with modern remote sensing technologies such as ENVI enable these objectives to come to fruition. Based on the coverage area, accuracy needs, and other project requirements, data variability and processing capabilities should be considered in advance of any project.

Data Variability

Many state and local governments already have aerial image collects over their AOI, and some recapture this coverage area every few years to detect and quantify changes over time. Many of these images have high spatial resolution – where one pixel on the computer screen represents a meter or less on the ground. Often times these data collects are 3-band RGB (red, green, blue) – also known as “true color” images where the collection device (sensor) is sensitive only to irradiated light in the visible portion of the electromagnetic spectrum. These images can be described as data with “high spatial and low spectral resolution”.

These types of data can often be processed very effectively when evaluating impervious surface coverage, and sometimes other landcover classes, as long as there is a fair spectral separation between materials. For example, the relationship between the data numbers (DNs) for a given pixel in each color channel can be plotted on a graph. The shapes of these plots are important characteristics for material identification where the shape itself can help to identify one material or another. Take the plots below where the blue plot represents a road pixel and the green plot represents vegetation. Note the differences between the shapes of the plots where the impervious surface plot has a fairly linear negative slope where red > g > b and the vegetation plot has a shape with a peak in band two where r < g > b.

The reason these relationships tend to hold true across the image is because of the high spatial resolution of the image. For example, if three-inch pixel data is available, there is a high probability that most of the pixels over a driveway contain only driveway reflectance. Conversely, if the spatial resolution were 15 meters per pixel, there would be a high probability that the majority of the driveway pixels might overlap with the lawn, dirt, trees, or other materials.

Also note that the band thresholds contain additional information that can be used to separate these materials where the band minimums and maximums, or data numbers themselves, fall within specific ranges in each of the color channels for a particular material. Note the value ranges of the plots below where the impervious surfaces appear to have higher DNs than those for vegetation.

Additional spectral information is often available as a fourth band in the near infrared (NIR) portion of the spectrum. When this band is available it is incredibly powerful for vegetation identification. This is because the physical structure of plants causes strong reflectance of NIR radiation. These high values are useful for not only separating vegetation from other green things, but to also classify things that are “not vegetation” thus improving classification results. The plot below represents a vegetation pixel in a four-band image.

Processes

Data Preprocessing

Before performing spectral processing it is ideal to perform some form of atmospheric correction to remove effects of atmospheric scattering. As you can see in the images and spectral plots below, the images may not appear to be very different; however the differences between the spectral plots before and after processing indicate that improvement of spectral separation can be achieved. These improvements are leveraged when using these spectral differences to differentiate materials.

Spectral Analysis and Supervised Classification

While approaches to generate impervious surface maps (ISM) can vary, utilizing available spectral information is a tried and true approach to qualifying pixels as those containing manmade vs. natural materials. Supervised classification considers in-scene pixels that are identified as belonging to one class or another (impervious and pervious in this case). These representative samples are usually selected from the scene and are known as end members. The shapes and values of the spectral plots – in other words – the spectral characteristics of these end members define reference base criteria.

Spectral classification algorithms evaluate each pixel and its relationship to each end member to determine the best class for the pixel. Different algorithms and variable parameters can produce variable results. ENVI’s Classification Workflow can be used in conjunction with the preview window to determine which algorithm gives the most accurate result for a particular scene. Alternatively, ENVI’s SPEAR TERCAT tool enables processing against many algorithms at once, and a Winner Take All approach considers the results from many algorithms.

Results Comparison

Common processing was performed on both the four-band and three-band NAIP images as specified above. Note that the same approaches elicit very similar results. The overall statistics of pervious vs. impervious surface coverage for the four-band image are 67.642% vs. 32.358% respectively. The overall statistics for the three-band image are 69.312% vs. 30.688% respectively.

Conclusion

Remote sensing technologies coupled with high spatial/low spectral resolution imagery provide a viable and realistic approach to create and quantify accurate impervious surface coverage maps. Low-cost or freely available imagery provide well-suited data options while also addressing budget constraints. Advanced classification approaches that include atmospheric correction and algorithms such as Spectral Angle Mapper and Winner Take All yield accurate maps and associated statistics. ENVI provides all of the tools necessary to ingest imagery, perform analysis, and fulfill project goals to generate GIS layers such as impervious surface coverage maps.