Sources of uncertainty

 
The ideal map would be accurate with a high level of precision and contain all the information that might be required by users. For example, fine-scale ordnance survey maps might be expected to show man-made objects in their correct positions with a very small margin of error. This is not the case for benthic habitat maps! Maps show the way the map-makers see the sea floor making best use of the data available to support their viewpoint. The following provides a description of some of the main causes of uncertainty:
 

Measurement error of the ground truth data

The natural environment is extremely complex and we need to simplify the real world considerably for mapping. The objects we map are usually our attempt to force the highly variable nature of the world into a manageable number of categories. It is inevitable that there will be ambiguity in this process which can originate from various sources:

  • Variability in the way surveyors apply a classification process to record data. The definitions of classes will be vague and many of the criteria will overlap from class to class. Error can be minimised by better definition of class attributes and standardised protocols for assigning samples to classes. Absence of clear guidelines makes it difficult and interpretation subjective.
  • Real heterogeneity on the ground. Variability is complex firstly because habitat features very often are on a continuum and lie between two or more habitat types in the classification, and secondly because fine-scale heterogeneity may result in the minimum mapping unit (MMU) or pixel encompassing more that one class.
  • Trying to fit observations limited by the technique used to a classification system where classes are based on more complete information. Video observations, for example, may not provide full information on infauna and the observation is classed on the basis of conspicuous fauna.

 

It cannot be stressed strongly enough that interpreting remotely sensed data using ground truth observations can be undermined by poor attribute measurement of the ground truth samples. This is particularly likely when the attributes are habitat classes and the analyst must decide how best to match the sample data to a classification system.

Subjective interpretation of boundaries

Many habitats are characterised by indiscrete or diffuse boundaries and are therefore subject to the interpretation or bias of the field mapper (for direct mapping) or visual interpretation of images (e.g. side scan images).

 

The inherent variability within and between the remote sensing systems

All remote sensing techniques have inherent variability that degrades their ability to discriminate features on the ground. Variability may also apply to distortion in video systems and the way different grabs of the same type ‘bite’ the sea floor, introducing observation error. Calibration of equipment is vital to the accuracy of the data and utilising poorly calibrated equipment will downgrade the accuracy of the final maps.

Positional errors of remote sensing, ground truthing and combined errors

The equipment we use will have limitations as to positional accuracy. Image processing requires the location of the ground truth samples on the image so that image characteristics can be associated with the ground truth classes. The combined positional errors will give rise to a locus (or ‘error envelope’). Thus, even if we could be absolutely precise about the mapping units, we could not be precise about where boundaries should be. Nor, because of discrimination, could we be absolutely sure we have detected the class with absolute certainty.

 

Error from analysis

Error and uncertainty will also be introduced through analysis, especially given that very often target classes themselves cannot be directly detected by remote techniques and their presence is inferred via statistical links to infauna or other observed variables (proxy maps, surrogacy). There can be many stages involved in image processing from data editing through to statistical analysis and modelling. However, the route followed by an analyst may be hard to replicate by another person since there are many possible pathways, each with different parameters that must be set. It is hoped that analysis is robust, but there is always the possibility that the interpretation is sensitive to apparently trivial parameter settings.

 

Error from sampling bias

Not every point in a map is validated. Maps are based on some form of sampling strategy and these data are extrapolated to the whole area using assumptions about the statistical relationship between the samples and the ‘population’ from which they are drawn. Wherever there is sampling there will be bias and problems of under-sampling. This is especially true for geographic systems where the uniqueness of location makes sampling strategy difficult.

 

Cartographic error

There is a limit to what a map can show (detail, number of classes and resolution) and maps generalise information to a greater or lesser extent. The ability to show detail in a map is determined by its scale. A scale of 1:2,000 will illustrate much finer points of data than a smaller scale map of 1:200,000. Scale restricts type, quantity, and quality of data. Enlarging a small scale map does not increase its level of accuracy or detail.

 
 
Links to websites:

All material variously copyrighted by MESH project partners 2004-2010

| List Access Keys | Mapping European Seabed Habitats (MESH) - Home | Overview | News | Work areas in the MESH Project | Partners | Contact Us | Search | Site Map | Partner Extranet |