Assessing the impacts of human uncertainty in the accuracy assessment of land-cover maps using linguistic scales and fuzzy intervals

• Published in: International journal of remote sensing 2015-05, Vol.36 (10), p.2524-2547
• Main Authors: Sarmento, Pedro, Fonte, Cidália C., Dinis, Joel, Stehman, Stephen V., Caetano, Mário
• Format: Article
• Language: English
• Subjects: Accuracy; Assessments; Case studies; Fuzzy; Fuzzy logic; Fuzzy set theory; Ground truth; Image resolution; Interpreters; Intervals; Land cover; Linguistics; Satellite imagery; Spaceborne remote sensing; Transformations; Uncertainty
• ISSN: 0143-1161; 1366-5901
• DOI: 10.1080/01431161.2015.1043034

The accuracy assessment of land-cover maps requires reference databases which are intended to represent ground truth. However, these reference databases are usually obtained through photo-interpretation of aerial or very-high-resolution satellite images and therefore have uncertainty which will influence the results of the accuracy assessment. Previous efforts to account for this source of uncertainty have employed a linguistic scale to translate the degree of correspondence between the ground conditions and each land-cover class for each sample location. The linguistic scale is transformed into fuzzy intervals with this transformation based on a photo-interpreter's hypothetical ideal perception of the land-cover areal coverage for a sample unit. The end result is a fuzzy accuracy assessment. The objectives of this article are to assess the degree to which the real response of photo-interpreters corresponds to the assumed ideal response and to evaluate the impact these differences have on the results of an accuracy assessment. To achieve this objective, we examine linguistic scales with five and seven values. Furthermore, we develop a method to transform these scales into interpreter-derived fuzzy intervals expressing the proportion of area of land cover for each sample unit. This transformation is accomplished using a control sample in which the area occupied by each land-cover class is assessed. The methodology is tested via a case study where a map with five land-cover classes is evaluated. The accuracy assessment is performed with both hypothetical ideal interpreter response and with the interpreter-derived fuzzy intervals. The results for the fuzzy accuracy measures produced from the different analyses show that there are considerable differences between the results obtained with the linguistic scale with five and seven values, and that the interpreter-derived seven-value linguistic scale provides results very similar to those obtained with the ideal interpreter response.