ASPRS POSITIONAL ACCURACY STANDARDS FOR DIGITAL GEOSPATIAL DATA

COMMENTS, NOTES AND BACKGROUND MATERIALS

(Page Last Updated: 1/8/2014)

1) COMMENT SUMMARY

Revision_Comment_History_Updated_0108_2015_Anonymous.xls

(Excel spreadsheet summarizing comments received and actions taken. The full version including original e-mails and links to supporting papers and information provided is available on request.)

2) KEY ISSUES FROM PAST MEETINGS, DISCUSSIONS AND DELIBERATION

This is not a comprehensive list of all issues discussed. The document provides detailed background information and discussion in support the assumptions made and final approaches selected therein. Additional background detail is also included in the first draft narrative version of the standard, as published in PE&RS in December 2013. The link to that document is posted on this web site. The discussion here is intended to provide background clarification for those selected key issues that received significant discussion and debate in the course of developing the draft standard.

Map scale, ground sample distance (GSD) and pixel size:There has been continued debate regarding how to assign accuracy classes to digital orthophotography and planimetric data,

Revision 6 presents accuracies that are independent of scale and pixel size. Associations to scale and pixel size, as related to both legacy data and current technologies, are now presented as guidelines in the Annex tables and not presented as standards.

Initial proposals utilized the GSD of the source imagery (ie. the distance on the ground represented by one pixel in the digital image). This was problematic since the accuracy of the final data depends on many factors other than the acquisition GSD. These include the sensor itself, ground control accuracies, direct georeferencing or aerotriangulation methods used, and compilation or rectification methods.

For digital orthophotography, a distinction was made between the GSD of the source image and the GSD of the orthophoto. GSD was used to refer to the raw image resolution; “pixel size” is used to refer to the resolution of the rectified orthophoto.

For planimetric data, the issue was more complicated. With digital data, there is no “published” map scale as the data can be printed or viewed at any zoom resolution. However, all planimetric data does have a specific resolution or target scale for which it was designed to support. This is determined by the resolution of the source imagery, compilation methodologies (ie. point spacing and the level of detail digizited) and other factors. While the data can certainly be viewed or printed at scales that exceed the intended target scale, doing so does not improve the accuracy or level of generalization. For this reason prior revisions based accuracies on a map scale factor which is defined as the design map scale, or the scale for which the data was designed to be viewed and printed. The Map Scale Factor was introduced to facilitate a simple formula for computing accuracy thresholds at any scale. This solution was workable for projects requiring scaled planimetric maps (whether digital or plotted), though still problematic for horizontal data not tied to scale or pixel size.

NSSDA equations in relating RMSE to 95% confidence levels:The NSSDA equations that use RMSE to compute the error range at a 95% confidence level are only statistically correct under the very restrictive condition that the mean error equals zero. This condition rarely occurs. Further, the RMSE statistics are only valid for normally distributed data. In the case of horizontal accuracies, the NSSDA equations only apply when the condition RMSEx=RMSEy is approached. However, in many common applications, empirical results indicate that if the mean error is small, and the data is normally distributed, the NSSDA equations can represent a reasonable approximation of the error range at a 95% confidence level. As the mean error increases, the approximation is less accurate and tends to overestimate the range of errors. This is not well documented in the NSSDA standard. In addition, RMSE by itself does not fully characterize other aspects of the data set accuracies. As such, RMSE is applicable only for normally distributed data sets with all systematic errors removed. Ensuring that the data set meets these requirements requires careful evaluation of the other statistical parameters.

Approach to Using RMSE in New Standard: RMSE is used in the existing standard and is a long established, well understood and widely used parameter for estimating geospatial data accuracies for normally distributed data sets. Further, as a simple to use, single parameter threshold, RMSE appropriately characterizes the absolute accuracy of the data set (as opposed to the relative accuracy or precision about the mean). For these reasons, the new standard continues to use RMSE as the accuracy threshold for normally distributed data sets. 95% values that correlate to the NSSDA reporting standard are listed in the table as reference. Clarifying text was added to indicate the limitations of that relationship and to explain the necessity of evaluating other statistical parameters to ensure that the data set has had all systematic errors removed and meets conditions for normally distributed data. Further, for lidar accuracies in vegetated areas (which are known to be biased and not normally distributed), thresholds are based on the 95th percentile accuracies and do not use RMSE values. The current standard was intended to provide simple to use and straight-forward thresholds for the most common data sets. It does not preclude future modules or addendum that address the more complex case of data sets that do not meet these criteria. In fact all efforts were made to facilitate incorporating this work when/if it is pursued.

Modular Standard: Several comments to date have indicated that other modules may be needed. These include: Assessment of linear data; rigorous total propagated uncertainty (TPU) models for our products (as opposed to ground truthing against independent data sources); more detailed statistics that do not rely on the assumption of normally distributed data; and image quality factors (such as edge definition and other characteristics). The current standard is intended to be the base standard needed to replace the existing standard for Large Scale Maps and to meet the immediate need of better addressing current digital technologies. Additional modules should be pursued and can be added by subject matter experts in these fields as they are developed.

Published data set vs. source data points:The standard is replacing the existing “map accuracy standard” and as such applies to values interpolated from the final data sets. The standard is not necessarily evaluating the system accuracy at discrete source points such as lidar returns or digitized points. Elevation accuracies are assessed as interpolated from a TIN generated from the final digital elevation model. Planimetric accuracies are measured at well-defined and readily identifiable features.

Spot Elevations and Contours:Higher accuracy spot elevation points are not specified by the new standard. Spot elevations were primarily used on cartographic contour maps, published at a fixed map scale, to aid in the accurate interpolation of elevations at key locations between drawn or interpolated contours. With current GIS, DTM and lidar technologies spot elevation points tied to a specific contour interval are less relevant. The new standard uses a single accuracy threshold to specify the accuracy of elevations interpolated from the source terrain model and moves away from specifying accuracies in terms of what can be interpolated from a published contour map. Since it is a derived product affected by interpolation errors and other factors, it should be noted that when tested or evaluated at check points, a generated contour map may not always achieve the same RMSEz accuracy as the source data.

More stringent that past standards:New technologies can achieve higher accuracies than required by prior standards. This standard facilitates more stringent accuracies than the existing 1990 standard.

Early comments indicated that, particularly for orthophoto imagery, additional classes may need to be added, or some of the higher classes may need to be relaxed to address applications where less stringent accuracies are required. Significant objections and concerns were raised with regards to proposed structure used in Revision 3 of the standard (March 2014) presented at the Spring Conference. That version listed Class 1 as the highest accuracy class appropriate for new technologies and very stringent project design; Class 2 corresponded to a typical high accuracy mapping project and corresponded to Class 1 in the old standard. The concern was that agencies and users would have difficulty adjusting to specifying Class 2 for the majority of their work, when Class 1 was the standard for so many years. To address this, Revision 5 introduced a Class 0 accuracy. In that version (Revision 5, July 2014), Class 0 represents the highest accuracy class; Class 1 corresponds to the Class 1 standard from the prior standard which is in widespread use and is still applicable for most high accuracy mapping projects.

Revision 6 solves this issue by abandoning discrete classes and referencing accuracy classes to the RMSEx and RMSEy thresholds they represent. This provides 100% flexibility for users and data providers to establish the most appropriate accuracy class for the application, given the very wide variation that now occurs in sensors, approaches and other critical factors. Annex B tables provide guidelines to choose the appropriate RMSE accurcy class as associated to both legacy data and prior standards as well as for what can be achieved with new technology.