Although spatial data users may not be aware of the inherent uncertainty in all the datasets they use, it is critical to evaluate data quality in order to understand the validity and limitations of any conclusions based on spatial data. Spatial data uncertainty is inevitable as all representations of the real world are imperfect. This topic presents the importance of understanding spatial data uncertainty and discusses major methods and models to communicate, represent, and quantify positional and attribute uncertainty in spatial data, including both analytical and simulation approaches. Geo-semantic uncertainty that involves vague geographic concepts and classes is also addressed from the perspectives of fuzzy-set approaches and cognitive experiments. Potential methods that can be implemented to assess the quality of large volumes of crowd-sourced geographic data are also discussed. Finally, this topic ends with future directions to further research on spatial data quality and uncertainty.
- Uncertainty modeling, simulation, and visualization
- Geo-semantic uncertainty
- Uncertainty in big geospatial data
- Future directions
A range of approaches have been developed to study both positional and attribute uncertainty. Positional uncertainty has been evaluated using analytical approaches and simulations. Analytical methods use mathematic formulae and statistical principles to describe systematic and random errors. For vector datasets that include point, polyline, and polygon, all analytical approaches are based on error representation of coordinate pairs. Error ellipse model and its derivatives have been developed to represent uncertain positions of a point using the variances in the x and y directions and the covariance of x and y coordinates (Alesheikh et al., 1999), with the center of the ellipse being the most likely position of a point. To represent uncertainty of polylines, different models have been developed based on point uncertainty models, including the epsilon-band model, the error band model, and the G-band model (Perkal, 1966; Dutton, 1992; Shi and Liu, 2000). The general idea of polyline uncertainty is a buffer around linear features, which is called the epsilon band. The buffer size of an epsilon band is dependent on factors including positional uncertainty of each node and spatial autocorrelation of points along a line. Positional uncertainty of polygons is usually caused by uncertainties of nodes and lines that compose a polygon, so it can be assessed using similar methods mentioned above. Furthermore, methods have been developed to evaluate characteristics of polygons, such as area, perimeter, and centroid. For instance, uncertainty of polygon area has been studied using the variances of its vertices and covariance between them (Chrisman and Yandell, 1988; Zhang and Kirby, 2000).
Analytical methods for representing positional uncertainty is rigorous and supported by statistical theories; however, the assumption of independent uncertainties of all vertices is rarely the case. Moreover, mathematical formulae can quickly become too complex to implement with complicated feature shapes and may not be the best way to communicate uncertainty with general data users. Another way to represent positional uncertainty is through simulation that generates a large number of realizations of geographic positions based on random variables. The most widely adopted simulation method is Monte Carlo simulation that involves several steps (Alesheikh, 1998): find the probability density function of errors in the input data, generate a set of random variables drawn from the probability density function, perform spatial operations to generate N output realizations of the random variables, and calculate summery statistics from the N realizations. For example, Monte Carlo simulation can be used to evaluate propagation of uncertainties of slope and aspect that are calculated from digital elevation models (DEMs).
Geo-semantic uncertainty generally refers to ambiguities and vagueness associated with geographic concepts, classes, and values that are sometimes subjective without agreed definitions. Examples of such concepts include hill and mountain (what elevation threshold distinguishes a hill from a mountain), downtown (where is the boundary of downtown Long Beach), and exurbanization (more than eighteen definitions are available according to Berube et al., 2006). It is not straightforward to measure this type of uncertainty using statistics or mathematic formulae. Geo-semantic uncertainty is usually studied using fuzzy-set approaches and cognitive experiments. Fuzzy sets allow an element to partially belong to a set with the aid of a membership function ranging from zero to one. For example, a set of membership functions of distances can be constructed to represent and visualize the ambiguities associated with the concept of exurbanization (Ban and Ahlqvist, 2009). As a result, various degrees of grayness may be used to represent the likelihood of a particular location being exurbanized or not exurbanized, which cannot be effectively visualized as crisp boundaries (Figure 2).
Figure 2. Uncertain boundaries of exurbanization using a fuzzy-set membership function. Darker color indicates higher degree of exurbanization (Li et al., 2018).
In addition to fuzzy-set approaches to study vague geographic concepts, behavioral and cognitive methods have also been adopted to visualize and evaluate geo-semantic uncertainty. This line of research relies on surveys and interviews to empirically determine the content of a vague spatial concept such as the extent of a region or spatial relations such as far and near. For example, Montello et al. (2003) recruited 36 participants to draw the outline of downtown Santa Barbara on a base map and aggregated these maps to visualize vague boundaries using dot-density shading. A more recent study by him and his colleagues examined the vague regions of “Northern” and “Southern” California using a similar approach (Montello et al., 2014). Meanwhile, there is a trend to use crowd-sourced geographic data to complement formal questionnaires and interviews to elicit people’s perception of vague spatial concepts. Li and Goodchild (2012) identified vague boundaries of several major places in France using geotagged photos collected from Flickr. Gao et al. (2017) replicated the study of northern and southern California regions using data from five sources of user-generated content and obtained similar results.
Uncertainty has always been a fundamental issue in GIScience. However, it is of special importance to assess uncertainty in big geospatial data, sometimes referred to as Volunteered Geographic Information (VGI, Goodchild, 2007), due to lack of quality control in data creation. Goodchild and Li (2012) proposed three methods to evaluate the quality of VGI. First, based on the crowd-sourcing approach, higher data accuracy is associated with a larger number of reviewers and editors. The effectiveness of this approach was confirmed by several studies on data quality of OpenStreetMap (e.g., Haklay, 2010). This method works best for areas or geographic features that attract interest from a large number of people, such as street networks in London. Second, the social approach establishes a hierarchy of data quality experts based on their performance and reputation in a particular data collection project. This approach mimics the system of social hierarchy in any profession, which arranges people in a relatively linear ladder. People at the top have more power and control over data management, especially in the case of a dispute. This approach has been adopted by many open and collaborative projects such as Wikipedia and OpenStreetMap. Although every user can create and edit any feature, people at a higher level make a decision regarding the quality of a particular piece of geographic information when a disagreement happens. Third, the geographic approach relies on geographic theories and principles to assess the quality of geographic data. A geographic fact inconsistent with geographic rules may be flagged as potentially inaccurate or incorrect. An example would be a bar mistakenly geotagged to a historical site location. Furthermore, open source geographic data may be compared with an authoritative or high-accuracy reference dataset to generate a quality control report, as demonstrated by Haklay (2010). One major limitation of this method is unavailability of a gold standard database as crowd-sourced geographic data cover a wide range of geographic themes not included in any official datasets.
Uncertainty is undoubtedly an essential characteristic of any geographic datasets and affects spatial analysis and decision-making. Researchers and scientists have made many efforts to understand, identify, evaluate, quantify, and reduce spatial data uncertainty and to increase spatial data quality using both analytical and simulation approaches. Although significant advances have been made in this area since the 1980s, there are still many challenges to be overcome. Many models have been developed to either mathematically or empirically assess positional accuracy while approaches for evaluating attribute uncertainty are still few, let alone models that assess both positional and attribute uncertainty coherently. Currently separate models are usually adopted to measure uncertainty in vector and raster datasets. In a real-world scenario, one application may require both vector and raster data in a single step. How to accurately assess uncertainty propagation that incorporates both data structures requires more work. In the era of big data, large amounts of geographic data may present a great opportunity to study physical environment and human society; however, new methods for understanding spatial data quality and uncertainty in crowd sources are to be developed and implemented. Rich geographic datasets are created on places without accurate or explicit positional information, novel computational and empirical approaches need to be discovered. Finally, although many models have been developed to investigate and communicate spatial data quality and uncertainty, they are not integrated into GIS software packages. Educating GIS professionals and geospatial technology users on the importance of spatial data quality and uncertainty still has a long way to go.
- Define spatial data uncertainty, quality, accuracy, error, and geo-semantic uncertainty
- Explain the importance of spatial data uncertainty
- Describe major methods to evaluate positional and attribute uncertainty
- Describe major methods to evaluate geo-semantic uncertainty
- Discuss possible methods to evaluate uncertainty of crowd-sourced geographic data
- Why is it important to understand uncertainty in geographic data?
- What are the two major categories of approaches to quantify positional uncertainty? What is an error ellipse model? What is the role of Monte Carlo simulation in studying spatial data uncertainty? What are the major steps of Monte Carlo simulation?
- What is an error matrix? How is it used in assessing attribute uncertainty?
- What are the two major lines of research in examining vague geographic concepts, such as downtown? What is a fuzzy set? How is it used in studying geo-semantic uncertainty?
- What are the four possible methods for evaluating quality of crowd-sourced geographic data?