National Technical University of Athens (NTUA)
Διερεύνηση οντολογιών, μηχανικής μάθησης, μορφομετρίας, και αντικειμενοστρεφούς ανάλυσης εικόνας για την αναγνώριση γεωμορφών, κτιρίων, και μεταβολών κτιρίων
Abstract
dc:descriptionThe objective of this research was the investigation, design, implementation, andenhancement of state-of-the-art computer vision and artificial intelligence approaches within the framework of GEographic Object-Based Image Analysis (GEOBIA) and their employment in specific remote sensing problems such as building extraction, foreshore mapping, building change detection and alluvial fan and bajada identification.The first objective of this dissertation was the design and implementation of an ontological reasoner allowing to design ontologies based on fuzzy spectral, geometric, and spatial relationships through multi-scale GEOBIA analysis, named SPatial Ontology Reasoner (SPOR). Identification of complex landscape components in terms of spectral, geometric, and rich spatial relationships, requires the representation of expert knowledge into a problem-solving strategy through an establish-and-refine-paradigm, within the environment of a knowledgerepresentation system. To take advantage of this expert knowledge within an automated image analysis system it is required to be formalized into a computer-conceivable form. Thus, a semantic gap arises between the high-level semantics employed by the experts to describe the phenomenon and the numerical low – level information extracted from data. To this end ontologies offer potential for knowledge formalization. Previous studies depicted the applicability of ontologies in GEOBIA. However, still remained the need to incorporate fuzzy reasoning in an ontology-based GEOBIA approach. Furthermore, the importance of spatialrelationships and multi-scale analysis has already been stated in earlier GEOBIA studies. Although, spatial reasoning with ontologies for image analysis has been examined, still remained the need for spatial reasoning in a multi-scale GEOBIA approach. Finally a limitation was stated for current Description Logic (DL) – based ontology reasoners, concerning the required processing time, when dealing with large number of classes and objects, which is very common in GEOBIA studies. Thus to address these needs in SPOR, OWL 2 was selected to formalize the ontology, which is latest version of this W3C standard. OWL 2 ensures the integration and exchange of GEOBIA ontologies with current semantic web technologies. Since GEOBIA frameworks should be compatible with Open GIS Standards, and it was also requiredcomputation of spatial relationships during the reasoning process, SPOR integratedPostgreSQL which supported such functionality.The second objective was the integration of advanced machine learning techniques within the ontological-based reasoning. In the literature machine learning techniques have been employed as a classification step before the ontology-based refinement in medical and scene analysis studies. However, still remains the need to examine the applicability of such approach in GEOBIA. To address this issue, it was decided to examine Deep Learning methods, which are machine learning algorithms aiming to model high-level abstractions in data by using model architectures composed of multiple non-linear transformations. Such algorithm is the Deep Belief Network. DBN training is performed in two phases. At first unsupervised training is performed by stacking multiple layers of simple unsupervised networks such as the Restricted Boltzmann machines (RBM) which were trained by the contrastive divergencealgorithm. Unsupervised training output is provided as input to Logistic Regression supervised algorithm which provides the final DBN output. To perform integrated reasoning with DBN and fuzzy ontologies, OWL 2 was enhanced with proper restrictions which allowed class definitions through the combination of DBN and fuzzy expressions. SPOR was also enhanced to be able to perform reasoning with such ontologies. The third objective was the investigation and addressing of specific remote sensing problems through GEOBIA multiscale analysis. To evaluate SPOR a building extraction approach through multi-scale analysis from QuickBird imagery was designed. Three segmentation and classification levels were designed and implemented. Aim of level 1 was the extraction of spectral categories with spectral properties similar to the present buildings. Aim of level 3 was the elimination of the road network. The final buildings were extracted on level 2. Evaluation with photo-interpretation data was satisfactory since 75% of the total rooftop area was extracted while the commission error was around 20%.Afterwards, an ontology-based multi-scale GEOBIA approach based on anisotropic diffusion and fuzzy ontologies for foreshore automatic identification was designed and implemented. The identification was based on the foreshore interpretation criteria provided by the Greek Cadastral Office, from multispectral and Digital Terrain Model (DTM) data. Management of the coastal zone is now recognized as an issue of importance due to the growing social and demographic pressures that threaten its sustainability. More than half of the world's population lives within 60 km of the coast. Specifically, foreshore mapping is important as the processes that take place in the foreshore affect the nearby water quality. Proper management of the foreshore will improve property values and provide recreational amenity. Foreshore mapping in Greece will contribute to the completion of the National Cadastre. Although in Greece this mapping was performed manually, it was decided to investigate methods for the formalization of the foreshore interpretation criteria provided by the Greek Cadastral Office towards the automation of the process. Due to the expression of the spatialorganization of the foreshore within the interpretation criteria (e.g. present vegetation, slopes, present infrastructure etc.) it was decided to formalize the criteria through an Ontology-Based GEOBIA approach. Due to the number and complexity of these criteria, the two most commonly appearing criteria in Greece were investigated, the vegetation border and the topographic crown border. To reduce undesired spectral and spatial detail, the imagery was filtered through morphological leveling which smoothed small spectral differences while major edges of the original image objects were preserved. A three level segmentation and classification hierarchy was developed were smaller objects were created on the lower levels, while larger objects where created on the higher levels. Level creation sequence was 1-2-3. On Level 1, general thematic categories were extracted (e.g. vegetation, water, low sloping areas). Through the interpretation rules derived from foreshore criteria these areas were refined on Level 2 and the final foreshore area was determined on Level 3.Evaluation with image interpreted reference data was satisfactory since 89% of the total foreshore area was identified while commission errors was around 14%.A building change detection scenario was investigated in suburban areas of Greece. The developed multi-scale GEOBIA approach integrated Deep Learning classification with ontological reasoning. A three level segmentation hierarchy was designed and implemented. Level creation order was 1-2-3. On the lowest level an estimation of areas of possible change was performed through fuzzy properties and multiple DBN classifications. Through interpretation rules related to the geometry of the buildings, employed on Level 2 and level 3 the final changes were derived. Results were satisfactory since 73 out of 87 changes were successfully identified, while having small omission and commission errors (10 and 14 changes respectively). To perform alluvial fan and bajada mapping, state-of-the-art DEM processing and multi-scaleGEOBIA analysis approaches were investigated, extended, and implemented. Mapping of alluvial fans and bajadas is important for practical and economic importance to society, particularly in arid and semiarid climates where they may be the principal groundwater source for irrigation farming and the sustenance of life. In this study, advanced image processing and knowledge representation methods were investigated and enhanced, towards the automatic identification of alluvial fans and bajadas. The developed method was based on the landform-pattern element approach which identifies landforms based on pattern-elements such astopography, drainage pattern and texture, soil spectral signature, spatial relationships, and other characteristics. Data used included the 10m spatial resolution National Elevation Dataset, provided by USGS and a pan-sharpened Landsat OLi imagery. The study area was located in the Death Valley, Nevada, USA. DTM processing included noise-removal filtering and depression treatment. A two-phase GEographic Object-Based Image Analysis (GEOBIA) approach involving multi-scale segmentation and fuzzy ontology-based reasoning was designed. Aim of the first phase was the identification of topographic forms such as the basin, mountain-range, and piedmont. Piedmont is the area where alluvial fans and bajadas reside.A four level segmentation hierarchy, accompanied by a four level ontological hierarchy, were developed. The identification was based on topography attributes such as slope gradient and curvature, and spatial reasoning. Aim of the second phase was the identification of alluvial fans and bajadas based on drainage pattern texture and topographic features such as landform shape. Thus a two-level segmentation and ontological hierarchy were developed. Due to a systematic omission on the fan toe, each fan and bajada soil spectral signature was taken into account to determine the omitted areas through region growing on the multispectral imagery. Accuracy assessment indicated satisfactory results, since a 90% completeness and 89% quality was achieved for both alluvial fans and bajadas.Finally, the fourth objective of this dissertation was the development and release of all the developed methods as Free and Open Source Software.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2017
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Argyridis, Argyros
- Αργυρίδης, Αργυρός
Subjects
dc:subject × 20- Άντικειμενοστρεφής ανάλυση εικόνας
- Ανίχνευση μεταβολών
- Κτίρια
- Εξαγωγή αιγιαλού
- Αλλουβιακά ριπίδια
- GEOBIA
- Ontologies
- Image analysis
- Buildings
- Change detection
- foreshore
- Alluvial fan
- Φυσικές Επιστήμες
- Γεωεπιστήμες και Επιστήμες Περιβάλλοντος
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Natural Sciences
- Earth and Related Environmental Sciences
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
dc:rights- Statement dc:rights
-
- BY_NC_SA
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/42008