• No se han encontrado resultados

MatchMaps has several advantages. Firstly, MatchMaps combines geospatial data matching and ontology matching techniques to generate matches between spatial features from disparate geospatial datasets. The experimental results show that MatchMaps outperforms several other ontology matching tools (which do not make effective use of location information) for matching geospatial data. In addition, MatchMaps is the only tool which uses qualitative spatial reason- ing and description logic reasoning to verify consistency of matches. The ex- perimental results show that this validation procedure improves precision and recall of output matches.

The main limitation of MatchMaps is its spatial reasoning does not use dif- ferent levels of tolerance for spatial features of different sizes and types (such as buildings, roads, rivers and lakes) and does not make use of direction in- formation. In the experiments above, the level of tolerance σ = 20m, which is not always appropriate. For example, if a1, a2, b1, b2 are small shops of size

less than 5m squared represented in two datasets, sameAs(a1, a2) is correct,

d(a1, b1) = 5m, d(a2, b2) = 30m, then sameAs(b1, b2) is wrong, but cannot be de-

tected using σ = 20m, because N EAR(a1, b1), N EAR(a2, b2). MatchMaps may

fail to detect wrong matches even varying σ values for different spatial features. For example, if sameAs(a1, a2) is correct, b1is N EAR and to the south of a1, and

b2 is N EAR and to the north of a2, then sameAs(b1, b2) is wrong but cannot be

detected by the spatial reasoning of MatchMaps.

In experiments, MatchMaps is used to match buildings and places. Though MatchMaps can match geometries and spatial objects, it does not actually dis- tinguish buildings and premises within buildings. For example, a named build- ing and premises within it are both treated as spatial objects. However, this issue is very complicated and out of the main focus of this thesis.

Other limitations include that MatchMaps cannot generate ‘partOf-disjointness’ axioms automatically, does not make use of temporal information from input data for information update, and has not automated the process of using matches for information enrichment and update. The limitations of MatchMaps indicate possible further works which are explained in the next chapter.

Conclusion and Future Work

This chapter concludes this thesis by summarizing its contributions and indi- cating possible directions for future work.

12.1

Conclusion

As EuroGeographics’ President, Ingrid Vanden Berghe [Geospatial PR, 2014], says, ‘Europe’s National Mapping, Cadastral and Land Registry Authorities must adapt their activities to become geospatial information brokers if they are to continue to meet society’s expectations’. This indicates that national map- ping agencies need to collate data rather than just collect data in future, except for areas where only national mapping agencies are able to collect the data. The rapid development of crowd-sourced geospatial data has provided challenges and opportunities to national mapping agencies. Compared to authoritative data, crowd-sourced data often contains richer user-based information and re- flects real world changes more quickly at a much lower cost, but it can be less accurate and less structured. In order to use crowd-sourced and authoritative

geospatial data synergistically, it is essential to establish matches between spa- tial features from different datasets.

This thesis is on matching disparate geospatial datasets and validating matches using qualitative spatial logic. It is built upon the state-of-the-art literature on geospatial data matching, ontology matching and spatial logic. Existing spatial logics and formalisms such as RCC8 are not appropriate for reasoning about crowd-sourced data. To reason about and debug sameAs and partOf matches between spatial features from different geospatial datasets, especially crowd- sourced datasets, a series of new qualitative spatial logics, LNF, LNFS and LBPT, was developed in this thesis. Previously, no spatial logic was developed for this purpose. The main technical results in the thesis are the soundness, completeness, decidability and complexity theorems for the new spatial logics with respect to a metric space (Chapters 7-9): a sound and complete axiomati- sation is provided and corresponding theorems are proved for each logic; the LNF, LNFS and LBPT satisfiability problems are all shown to be NP-complete. The spatial logics are proved to be proper fragments of the logic M S(M ) de- scribed in Section3.3.3for reasoning about distances, but have lower computa- tional complexity (provided NP ( EXPTIME ). It is also proved that the LNF satisfiability problem is decidable in PSPACE with respect to a 2D Euclidean space in Chapter7, whilst the M S(M ) satisfiability problem in a 2D Euclidean space is undecidable.

Another contribution of the thesis is a software tool MatchMaps (described in Chapter 4) for generating and validating matches between spatial features from different geospatial datasets. It generates candidate matches using loca- tion and lexical information (Chapter5), and validates matches using reasoning in description logic and the qualitative spatial logic LBPT (Chapter6and Chap- ter10). Previously, no matching tool used spatial logic for validating matches. Description logic has been used in several ontology matching systems, but not

for reasoning about crowd-sourced geospatial data. MatchMaps is the only sys- tem which combines spatial logic reasoning and description logic reasoning to check the consistency of matches with respect to location information and clas- sification information. Description logic is used to validate matches regard- ing classification information, more specifically, to exclude sameAs matches be- tween spatial features in disjoint categories or classifications, and also to reason about for which types of spatial features partOf matches cannot hold. Quali- tative spatial logic is applied to validate matches regarding location informa- tion. For example, for spatial features a1, b1 in one dataset and a2, b2 in another

dataset, a contradiction arises if sameAs(a1, a2), sameAs(b1, b2), NEAR(a1, b1),

FAR(a2, b2). As NEAR and FAR statements are treated as facts, at least one

of the sameAs matches is wrong and should be removed to restore consistency. MatchMaps involves domain experts in the process of validating matches to de- cide the correctness of such matches (which cause a contradiction) and remove incorrect ones, as no heuristic for making such decisions automatically gives sufficiently reliable results. The performance of MatchMaps was evaluated by the author and experts from Ordnance Survey of Great Britain (Chapter 11). Experimental results show that MatchMaps achieved high precision and recall, as well as reduced human effort. MatchMaps outperformed several ontology matching systems mainly because they cannot make effective use of location information.

Though the work presented in this thesis is motivated by the development of crowd-sourced geospatial data and aims to use crowd-sourced and authori- tative geospatial data synergistically, the methodology developed and imple- mented in MatchMaps has wider applications in matching geospatial datasets containing vector data, not limited to a crowd-sourced dataset and an authori- tative dataset.

Documento similar