IEEE Geoscience and Remote Sensing Magazine - September 2016 - 32

corresponding ontologies are given as input to GeoTriples,
which automatically constructs R2RML or RML mappings that
dictate the method of conversion of data into the RDF data
model. Spatial information is mapped into RDF according to
the GeoSPARQL vocabulary by using the transformation extension functions for R2RML/RML developed in LEO.
STORAgE/qUERyINg
MonetDB and Strabon are used as storage and query engines
in this application. As we explained earlier, MonetDB is used
for the storage of raw image data and metadata, handling
them as arrays and processing them using SciQL queries. In
the "Transformation into RDF" section, we also explained
how the hot-spot products are encoded in stRDF, so that they
can be combined with auxiliary linked geospatial data. By
correlating detected hot spots with auxiliary data, we associate more sophisticated information with them (e.g., municipalities that are affected by the fire), and we increase their accuracy (e.g., false alarms are detected and discarded). These
rich queries can be expressed by exploiting the expressive
power of the query language stSPARQL.
qUALITy ASSURANCE
The data sets described previously are mainly used to enhance the information captured by hot-spot data and to
increase its accuracy. In this section, we describe a series of
refinement steps using stSPARQL updates that enrich hotspot data with information about nearby municipalities
and increase its accuracy by detecting and correcting false
positives or omission errors.
Notably, the queries described below are sophisticated
update statements that exploit extensively the expressivity
of SPARQL 1.1 and stSPARQL (e.g., GROUP BY, HAVING,
aggregations, and OPTIONAL) to cover the needs of the real-time wildfire-monitoring application of the NOA. GeoSPARQL does not support spatial aggregates and updates
that were needed in this use case.
attribute enrichment
Each hot spot is connected with a municipality where it is
located, using the GAG data set. This is crucial information
offered to decision makers and crisis managers for the optimal allocation of their firefighting resources. In the following statement, the name of a municipality that spatially
intersects with a hot spot is set as a property of the hot spot
to perform this refinement.
INSERT {?h gag:hasMunicipality ?municipality}
WHERE {SELECT ?h (SAMPLE(?mLabel) AS ?municipality)
WHERE {?h rdf:type noa:Hotspot;
noa:hasGeometry ?hGeo;
?m rdf:type gag:Municipality;
rdfs:label ?mLabel;
strdf:hasGeometry ?mGeo.
FILTER(strdf:intersects(?hGeo,
?mGeo)).} GROUP BY ?h}
32

consistency refinement
The thematic consistency of the hot spots generated by the
processing chain is achieved by the refinement step that correlates them with auxiliary geospatial data. This is done by a
series of stSPARQL updates on the stRDF representation of the
hot spots by taking into account relevant stRDF data sets from
the ones presented above. The first step is to delete all hot spots
that lie in the sea. Classification of the pixels inside the sea
as hot spots is commonly encountered in wildfire scenarios
near the coast. The hot smoke spreads above the sea, leading to
misclassification of the corresponding pixels. This operation
is performed by the following stSPARQL update that marks
every retrieved hot spot as discarded. We are using the query
applied previously, i.e., if a hot spot does not spatially intersect
a municipality, it lies in the sea; so it is discarded.
INSERT {?h noa:isDiscarded "1"^^xsd:int}
WHERE {?h rdf:type noa:Hotspot.
OPTIONAL {?h gag:hasMunicipality?
municipality}.
FILTER(!bound(?municipality)).}
In a similar way, hot spots that are in the mainland but lie
in nonconsistent land cover areas are also discarded. Using the
Greek Landscape data set, the nonconsistent classes are defined as artificial surfaces, agricultural areas (i.e., arable land
and permanent crops), and wetlands and water bodies. This
operation is performed by the following stSPARQL update.
INSERT {?h noa:isDiscarded "1"^^xsd:int}
WHERE {SELECT ?h
WHERE {
?h rdf:type noa:Hotspot; noa:
hasGeometry ?hGeo.
?a rdf:type clc:ExcludeArea;
clc:hasGeometry ?aGeo.
FILTER(strdf:mbbIntersects(?hGeo,
?aGeo)).}
GROUP BY ?h ?hGeo
HAVING strdf:contains(strdf:union
(?aGeo),?hGeo)}
To ensure the product visualization consistency, we also utilize
the Greek Coastline data set, keep only the part of a hot-spot
polygon that lies in land, and eliminate the part that lies in the sea.
temporaL persistence
The fire-detecting processing chain identifies hot-spot pixels and marks them either as potential fires with a confidence
level of 0.5 or as certain fires with a confidence level 1.0. The
algorithm is based on a series of spectral tests with some
thresholds. Appropriately setting these thresholds is the outcome of a tradeoff between omission errors and false alarms.
In certain scenarios, this leads to a phenomenon described as
the Christmas tree effect, where some hot spots appear for the
first time, in the next timestamp they disappear, then they
ieee Geoscience and remote sensinG maGazine

september 2016

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2016