With the amount of textual data available on the web, new methodologies of knowledge extraction domain are provided. Some original methods allow the users to combine different types of data in order to extract relevant information. In this context, this paper draws the main objectives of the ANIMITEX project which combines spatial and textual data. The data preprocessing step is detailed.
A lot of high resolution satellite data are now available. This raises the issue of fast and effective satellite image analysis as it still requires a costly human implication. In this context, remote sensing approaches enable to tackle this challenge. The exploratory and ambitious ANIMITEX project1 aims at processing massive and heterogeneous textual data (i.e. big data context) in order to provide crucial information to enrich the analysis of satellite images.
The large amount of data are associated to a temporal repetitivity that increases. For instance today around ten images are available per year (e.g. SPOT, Landsat), and in 3 years, one image every 5 days (based on Sentinel-2 satellites) will be available.
1http://www.lirmm.fr/⇠ mroche/ANIMITEX/ (web site in French)
The ANIMITEX project has many application
areas such as image annotation (Forestier et al. 2010).
For instance, identifying the precise type of culture
or the function of a building is not always
possible with the only use of images. Nevertheless,
textual data could contain this kind of information and
give additional meaning to the images. The
development of approaches based on image/text matching
becomes crucial in order to complete image
analysis tasks
Moreover, image-text matching will enrich
Information Retrieval (IR) methods and it will provide
users a more global context of data
In the context of the ANIMITEX project, we plan to investigate two specific scenarios: (i) The construction of a road on the north of Villeveyrac (city close to Montpellier, France), (ii) A port activity area, called Hinterland, in Thau area (near to Se`te, France). The aim of this case studies is to enrich images with information present in documents, e.g. the opinions extracted in newspapers about land-use planning.
The section 2 describes the proposed data preprocessing step. The section 3 details the partners involved in the project.
The current work focuses on adapting of Natural
Language Processing (NLP) techniques for
recognition of Spatial Features (SF) and thematic/temporal
information
A detailed classification of the land occupation is currently in progress. It will lead to a digital vector layer where each SF (represented by a polygon) belongs to a class of specific land use. The nomenclature of this classification is organized into four hierarchical levels (See Figure 2). Moreover we investigate multi-scale information associated with different levels of classification of satellite images.
From this corpus, NLP methods have been
applied in order to identify linguistic features
concerning spatial, thematic, and temporal information
in the documents. The combined use of lexicons
and dedicated rules
SF validation: for each identified ASF, we check on external resources if there is a corresponding spatial representation. In particular, we have used layers provided by the IGN3 (municipalities, roads, railways, buildings, etc.). In addition, if an ASF does not present on IGN ressources, we use gazetteers (Geonames and Open Street Maps) to complet the information. Concerning the representation of RSF, we use spatial indicators of topological order associates to ASF.
Following the scopes proposed in
For this, we use the centroid of ASF and we 3Institut National de l’information Gographique et forestire, i.e. National Institute of Geography build a buffer around. The size of this buffer will be calculated taking into account the surface of the studied ASF. Then we decompose the buffer into four equal areas (forming a ”X”) from the centroid. Each intersection between the buffer and cones thus formed represent the four cardinal points. • Geometry: geometry relations are built from at least two ASF. These relationships are, for example, the union, the adjacency, the difference or a position of an ASF with respect to other ASF, for example, C between A and B (where A,B and C are ASF), etc.
the extraction step and spatial representation of the ASF and RSF, the spatial footprint associated with the treated document can be mapped. In this process, two main problems have been identified. The first one is the persistent ambiguity of some NE contained in SF because of some NE (e.g. Montagnac) corresponding to several places. To address this issue, a configurable spatial filter based on predefined scenarios has been developed. For example, to identify events related to a specific land-use planning project occurred in a part of the area of the Thau lagoon, only the SF contained in this area will be explored. The second issue is related to the use of external resources and the identification of the spatial representation appropriate to each ASF. Taking into account the spatial indicator (e.g. town, road, etc.) preceding by the toponymic name is a first answer because it allows us to specify the type of the SF and thus take into account the appropriate spatial representation.
Thematic information is identified by semantic
resources (i.e. AGROVOC thesaurus, nomenclature
resulting of image classifications ...)
These linguistic features allow us to identify specific phenomena in documents (e.g., land-use planning, environmental change, natural disasters, etc.). The main idea is to link the phenomena identified in images with subjects found in documents during the same period. Overall, the ANIMITEX project allows the users to integrate different information sources, i.e. both types of expressions (texts vs. images). The main objective is to enrich the information conveyed by a text with images and vice versa. 3
The multidisciplinary consortium of the project involves three research domains: Computer Science, Geography and Remote Sensing. More precisely, the expertise in remote sensing and complex mining and heterogeneous spatio-temporal data, is one of the foundations of the project.
TETIS (Territories, Environment, Remote Sensing and Spatial Information, Montpellier) aims to produce and disseminate knowledge, concepts, methods, and tools to characterize and understand the dynamics of rural areas and territories, and control spatial information on these systems. LIRMM (Informatics, Robotics and Microelectronics, Montpellier) focuses on knowledge extraction. ICube (Strasbourg) is specialized in image analysis and complex data mining. ICube collaborates with geographers from LIVE laboratory (Image Laboratory, City, and Environment) and specialists in NLP (LiLPa lab – Linguistics, language, speech). These two locations (Montpellier and Strasbourg) constitute a cluster of local skills related to all major aspects of the project. LIUPPA (Pau) includes researchers specializing in Information Extraction (IE) and Information Retrieval (IR). The main work of this partner is about extraction and managment of geographical information. GREYC (Caen) brings researchers in data mining (e.g. mining sequences in order to discover relationships between spatial entities) and NLP. For this aspect, a collaborations with two other labs is developed (LIPN and IRISA).
The authors thank Midi Libre (French newspaper) for its expertise on the corpus and all the partners of ANIMITEX project for their involvement. This work is partially funded by Mastodons CNRS grant and GEOSUD Equipex.