LiDAR: An End User's Perspective

This video, LiDAR: An End User's Perspective, is the second presentation I gave at the New York City LiDAR workshop on August 2, 2010.  I posted a recording of my first presentation, LiDAR 101, a few days ago.  In this presentation I share some advice based on the experience my colleagues and I in the Spatial Analysis Lab have gained over the last six years working with LiDAR data.  Certainly not all encompassing, but I do hope that it provides some useful information.

The LiDAR point cloud visualizations, surface models, along with the various LiDAR-based analyses were performed using Quick Terrain (QT) Modeler.  At the end I demonstrate a workflow for automatically mapping land cover by fusing imagery, LiDAR products from QT Modeler, and vector datasets using eCognition.

Is the Peer Reviewed Literature the Best Place to Publish?

This New York Times article resurrected a question that pops up in my head every time I read a peer reviewed article in a remote sensing journal that involves some type of  land cover mapping: is this really the best medium for publishing this work?
The typical journal article with have an introduction, methods, results, and discussion.  There will be 1-2 screen captures of the final product, most likely in black and white, and then an accuracy assessment.  The screen captures are typically of a relatively small area, and often at a scale not suitable for assessing the quality of the data.  The accuracy assessment is of course done by the very people who did the analysis, which is hardly an independent assessment.  The problem is further exacerbated by the fact that the slow pace of peer review means that an article published today is likely the result of techniques applied 2, 3, or more years ago.  In the rapidly evolving remote sensing field this is a lifetime.  There are strong arguments for the peer review process, but if the reviewer does not have access to the end product they are simply reviewing a story about the data.

I am of the opinion that if some sort of final product, such as a land cover map, is the basis for the article, that dataset should be made available.  With all the advances in mash-ups and GIS servers it is not that difficult to do.  What of instead of publishing those papers to peer reviewed journals they were just posted to a blog, with an accompanying link to the map?  The article could be published as soon as the project is complete, as opposed to waiting years, and readers could view the final product and provide feedback.  This would allow others to replicate the techniques in a timely manner, which is particularly valuable given the ways in which a software package can change in a 1-2 year period.  It would also allow persons with expert knowledge of the area, particularly decision makers who might make use of the data, to assess its quality.

You will notice that I did not use the term "accuracy."  "Quality,"  while subjective, is a superior measure for assessing land cover maps, particularly those derived from high-resolution land cover data.  Quality implies not only accuracy, but a realistic representation of the landscape.  The distinction is illustrated quite well in the graphic presented below.  This datasets (not mine) passed peer review where it was purported to have a near 90% overall accuracy.  The screen capture shows three land cover classes (tree canopy, impervious, and grass/shrub) with the legend purposely omitted and the colors designed to be confusing.  The land cover is from a medium density urban area at a scale of 1:1,000.  While the data might meet some peer reviewed threshold for accuracy, I think it is safe to say it falls short on the quality front.  The simple fact is that for a residential area, if you have a map depicting tree canopy, grass, and impervious and the end user cannot make out what those classes are without a legend, you have a problem.  It is relatively easy to hide these shortcomings in peer reviewed article.  If one does not have to display the end result, there is an incentive to focus instead on a "fancy" methodology, throwing around terms like "neural networks." 

Below is an example  from one of our projects displayed at a scale of 1:1000, although from a different medium density residential area.   Once again, I am not displaying the legend, but the same colors are used to represent tree canopy, grass, and impervious.   I would be interested to hear feedback, but I bet you can make out the land cover classes based on the arrangement of the features.  Interestingly, the accuracy of this land cover data set is only slightly better than that of the one above.  The difference is more subjective, the difference is in the quality.

Did the Land Cover Change?

The City of Roanoke, Virginia released the following press release on their web site.

In 2002, an Urban Ecosystem Analysis found that Roanoke’s tree canopy was 32 percent, but since then the city has planted almost 3,000 new trees, with a goal of achieving 40 percent tree canopy in 10 years. In early 2010, a Virginia Department of Forestry report showed the city’s tree canopy to be 48 percent of total land area. The report also allows city forestry staff to target specific neighborhoods, blocks, or public properties that do not yet meet the 40 percent goal.

On the surface it would lead one to think that planting 3000 trees resulted in the overall tree canopy increasing from 32% to 48% (over 17 sq km) in a relatively short period of time.  This equates to 1.4 acres of tree canopy for each one of the 3000 trees planted.  Even if natural growth is factored in this rapid increase in canopy is impossible to achieve in a temperate ecosystem, particularly as some tree canopy would likely have been lost during the same time period.  The 2002 Urban Ecosystem Analysis was done by American Forests, and appears to be a pixel-based classification of high-resolution satellite imagery.  The 2010 report from the Virginia Department of Forestry was based on the 2008 1m NAIP, with object-based techniques used extract land cover.  The bottom line is that using land cover datasets derived from different sensors using different methodologies to report change yields misleading information.

Change detection, particularly high-resolution land cover change detection in heterogeneous urban areas, is extremely challenging, but it is something we as the remote sensing community will need to master if we are to provide decision makers with accurate information regarding the impact of policies and initiatives.  In the meantime the City of Roanoke could estimate the change in canopy from 2002 to 2008 by taking the two sets of images, dropping a few thousand sample points, and categorizing them and canopy or not canopy based on manual interpretation.  This technique tends to be quite accurate for producing city-wide estimates of change.  Unfortunately it would not help in determining where the change occurred.  This would require a more costly approach, generating consistent land cover datasets from both the 2002 and 2008 time periods.

LiDAR 101: NYC LiDAR Workshop

The New York City LiDAR acquisition this past spring received quite a bit of press.  In early August the New York City Urban Field Station and the Mayor's Office of Long Term Planning and Sustainability organized a workshop to introduce various city agencies to LiDAR in hopes of finding the best way to capitalize on this investment and to educate the end user community.  This video is a rerecording of a talk I gave, entitled "LiDAR 101."  I would like to thank Sean Ahearn from CARSI for providing me with one of the LiDAR tiles.