Robust Image Analytics and Crowdsourced Geographic Data
I've been reading a summary paper "Crowdsourced Geospatial Data" that I found on the Defense Technical Information Center website. The paper discusses a variety of crowdsource projects and common methods for quality assessment of the collected data. There are many great takeaways in the paper, but it drove me to think of how robust imagery analytics might be applied to improve the collection of this volunteered geographic information.
The authors point out that people participating in these projects are not always experts. They do have a passion for contributing, whether it be adding roads and other primary features in Open Street Map, or using Tomnod to delineate storm damage areas or locate shelters used by internally displaced persons (IDPs). In any volunteer effort, it is essential to keep the passion for participation high while maximizing the quality of the results. For crowdsourcing tasks that rely on information extraction from imagery, robust analytics can assist non-expert volunteers in a couple of ways.
First, in the case of a natural disaster, the task is often to delineate the boundary of the event. You might think that damage areas are easy to spot in imagery, and often, they are. However, when asking a person to spend hours or days doing this work, we should make it easy (keep passion high!). Automated change detection techniques applied to pre- and post-disaster imagery can highlight the damage areas. These techniques can provide overlays on the imagery that make it easier to locate and subsequently trace damage areas. And for neighborhood scale objects like buildings, robust change detection may find damaged objects that a volunteer could miss.
The image above shows an automated change detection overlay. The bright red pixels represent vegetation that was destroyed in a tsunami. Muted red pixels were impacted less and black pixels indicate no change. Image courtesy of Exelis.
Second, the search for shelters or other indicators of displaced persons may require viewing thousands of satellite images. This is time consuming and taxes the volunteer's eyes. There is a category of robust remote sensing analytics referred to as broad area search methods. These methods utilize the spectral content common in today's commercial satellite imagery to narrow in on objects of interest. Algorithms for anomaly detection, spectral complexity, and material identification are examples. Running these analytics on satellite imagery results in a map of pixels of where to look first. Again, this is an overlay that directs the volunteer to the parts of the image most likely to contain shelters (or other objects of interest) and often uncovers things that are invisible to strained eyes.
The image on the left shows a multispectral image with an anomaly detection overlay. The green pixels in the overlay represent features in the image that are unique against their background. In this desert environment, the anomalies correspond well with man made materials that could be shelters. The image on the right is a magnified view of a high resolution panchromatic image.