Multi-image movie of Pu’u ‘O’o Crater. (5/16-29/14)

Pu’u ‘O’o Crater East Flank. (5/16/14-5/29/14)

(Volcano Watch is a weekly article written by scientists at the U.S. Geological Survey’s Hawaiian Volcano Observatory.)

Hawai‘i County is the fastest-growing region in the State of Hawaii. Its relatively inexpensive real estate, along with population growth, spurs more and more construction on the flanks of its active volcanoes.

With growth comes increased risk from lava-flow hazards. Estimating the potential damage from future lava flows to infrastructure or development poses serious challenges to emergency managers.

A new tool that can help identify structures in a lava flow’s path is image classification applied to satellite imagery. This technique, combined with estimates of future lava-flow coverage, enables civil authorities to know the number of structures in any geographic region that are threatened by lava flows.

Image classification is widely used in agricultural studies to create land-use maps, but it has not been used extensively in volcanic hazards assessments. We are testing the applicability of the image classification technique––using remote sensing images, like those found on Google Earth––in combination with geographic information systems software. We attempted to automatically locate and classify structures using both supervised and unsupervised classification schemes.

In the supervised classification schemes, the user is able to select those areas in an image that correspond to forest, houses, or roads and use these classes to produce a classified image. In unsupervised classification, which is not dependent on the user, the algorithm auto-selects the features. We used both methods on different regions of the Big Island to estimate the numbers of structures present and checked their accuracy by manually counting the number of structures.

For two test areas where there is high risk from future lava flows—Hawaiian Ocean View Estates (HOVE) and Leilani Estates—the results of the supervised classification were more accurate than those of the unsupervised technique. The supervised yielded 85 percent for HOVE and 78 percent for Leilani, compared with 75 and 64 percent, respectively, for unsupervised image classifications.

For two other test cases, the unsupervised classifications yielded better results––92 and 95 percent––in detecting structures than the supervised classifications, which yielded 82 and 85 percent. Therefore, while it may be difficult to know in advance which classification scheme is more accurate than the other, the results give us a reasonable degree of confidence that these techniques could be useful.

The size of the area and the structure density appear to influence whether one classification scheme might be more accurate than the other. For example, when assessing a small area with a high structure density, the unsupervised classification appears to be more accurate than the supervised classification. On other hand, the supervised classification yields more accurate results in a larger area with relatively low structure density. We will explore the reasons for the differences in future tests.

The difference between the algorithms lies in the way the classes are grouped. For the supervised classification, the user defines each feature type or class beforehand and builds a signature file with this information. Different colors define different types of features, such as vegetation, roads, or houses. In unsupervised classification schemes, each pixel is compared with the color class or signature of the feature type. Then all the pixels in the study area are assigned automatically to a known class to which it has the highest probability of being a member.

Using supervised and unsupervised classification schemes for each assessment might combine the strengths of both methods. For example, we could use the unsupervised classification first to produce a signature file and then use that signature file to run the supervised classification.

We continue to refine our classification algorithms and further enhance them by eliminating false identifications. Our study shows that classifying satellite images can be useful in estimating the number of structures potentially in harm’s way. Ultimately, it will enhance our ability to assess our island’s vulnerability to future lava flows.

Map showing the Kahaualeʻa 2 flow in relation to the eastern part of the Island of Hawaiʻi as of May 30, 2014. Red stars mark active breakouts—the most distant was 6.2 km (3.9 miles) straight-line distance northeast of Puʻu ʻŌʻō. Older lava flows are distinguished by color: episodes 1–48b flows (1983–1986) are shown in gray; episodes 48c–49 flows (1986–1992) are pale yellow; episodes 50–55 flows (1992–2007) are tan; episodes 58–60 flows (2007–2011) are pale orange, and episode 61 flows (2011–2013) are reddish orange. The active lava tube is shown with a yellow line (dashed where its position is poorly known).