Abstract
Today satellite images are mostly exploited automatically due to advances in image classification methods. Manual visual image interpretation (MVII), however, still plays a significant role e.g., to generate training data for machine-learning algorithms or for validation purposes. In certain urban environments, however, of e.g., highest densities and structural complexity, textural and spectral complications in overlapping roof-structures still demand the human interpreter if one aims to capture individual building structures. The cognitive perception and real-world experience are still inevitable. Against these backgrounds, this article aims at quantifying and interpreting the uncertainties of mapping rooftop footprints of such areas. We focus on the agreement among interpreters and which aspects of perception and elements of image interpretation affect mapping. Ten test persons digitized six complex built-up areas. Hereby, we receive quantitative information about spatial variables of buildings to systematically check the consistency and congruence of results. An additional questionnaire reveals qualitative information about obstacles. Generally, we find large differences among interpreters’ mapping results and a high consistency of results for the same interpreter. We measure rising deviations correlate with a rising morphologic complexity. High degrees of individuality are expressed e.g., in time consumption, in-situ- or geographic information system (GIS)-precognition whereas data source mostly influences the mapping procedure. By this study, we aim to fill a gap as prior research using MVII often does not implement an uncertainty analysis or quantify mapping aberrations. We conclude that remote sensing studies should not only rely unquestioned on MVII for validation; furthermore, data and methods are needed to suspend uncertainty.
Highlights
I N THE field of geoinformation, “uncertainty” can be understood as a human “intrinsic” feature that “exists in the whole process from geographical abstraction, data acquisition, and geo-processing to the use of data” [1])
For the classified number of buildings, we find a significant coefficient of variations (CV) of 27.7% and with respect to building size we find a substantial CV of 20% among the interpreters (Table II)
We find that image interpretation elements as “key issue” of perception have an impact on the digitization
Summary
I N THE field of geoinformation, “uncertainty” can be understood as a human “intrinsic” feature that “exists in the whole process from geographical abstraction, data acquisition, and geo-processing to the use of data” [1]). For almost all quantitative performance evaluations, and especially to unveil the strength of machine learning algorithms, previous knowledge on spatial and thematic content of the data is of utmost importance. This previous knowledge is often – if it even exists – not consistent, for spatially highly resolved data and very specific types of land use. We find this gap in complex urban environments, which are characterized by multifaceted built-up structures with changing land uses in very close vicinity. The setting is based in very complex urban structures characterizing poor urban areas, e.g., slums/informal settlements and deprived formal areas
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
