What is accuracy?

Accuracy as applied to habitat mapping is a measure of the predictive power of a map to represent the world as measured against reality, while error is a measure of the departure of a map from reality. It is a mathematical measure based on 'hits and misses' (successful predictions and erroneous predictions). Note that this definition of accuracy is focused on the correct prediction of a habitat class at a particular location (in a vector map) or a pixel (in a raster map). In other words, there are two elements to accuracy: the correct class at the correct location. This definition is often termed classification accuracy; in other words have the data at point X been correctly classified?. Clearly, there is a positional element to this accuracy. For example, are boundaries between adjacent habitats accurately located? This could be restated as 'do the change in predicted habitats accurately mark the boundary between them in reality?'

Accuracy could be used as one of the criteria for assessing confidence. However, a strict mathematical measure of accuracy could be misleading, especially if two or more maps are being compared. For example, one map might class habitats in an area as either rocky or sandy and detail these two classes with a high level of accuracy. Another might show each of these habitats as a patchwork of different types of rocky or sandy habitats. The second is likely to be far less accurate but contain more useful information allowing for a certain level of error. The italicised phrase stresses the important point that some user-judgement has entered the assessment to make allowances for the lower accuracy. Thus, a user may have more confidence in the information contained in the second map despite its lower accuracy. The problem is that although many of the accuracy measures are mathematically sound, they still do not address the main issue of the overall confidence with which maps should be regarded. The same measure applied to different maps may give an erroneous impression of their relative 'success'.

Indeed, there is often a trade-off between information content and accuracy of a map: a map showing a large number of classes on a particular theme contains more information than one with a small number of classes. However, the error associated with the predicted distribution of the former might be quite high.

A schematic showing the changing relative importance of generalisation versus detail as the scale of a map changes.

All material variously copyrighted by MESH project partners 2004-2010