Ontology-based semantic classification of satellite images: Case of major disasters

Abstract

Abstract-The International Charter1 "Space and Major Disasters" is regularly activated during a catastrophic event and offers rescue team damage maps. Most of these maps are built by means of satellite image manual processing, which is often complex and demanding in terms of time and energy. Automatic processing offers prompt treatment; nevertheless it usually presents a semantic gap handicap. The exploitation of ontologies to bridge the semantic gap has been widely recommended due to their quality of knowledge representation, expression, and discovery. In this work, we present an ontology-based semantic hierarchical classification method to undertake this problem. Ontology components are translated to image-based parameters and used to assist the classification process with two levels and 12 embedded classes. The region of interest is selected from the first level, and exhaustively analyzed and classified at the second level. The 2010 Haiti earthquake was selected as study area for this work. Experiments were performed using very high resolution multi-temporal QuickBird imagery and eCognition software.

Geographic ontology for major disasters: Methodology and implementation

“…For future work, and as well as improving the ontology in terms of a comprehensive domain representation, we will be looking for its applications into other areas. Of particular interest is the ontology-based semantic annotation of satellite images [43], change detection, and response to queries related to emergency needs, such as the location of hospitals compared to affected areas, detecting operational roads, and location of the highest priority areas (schools, residential buildings, etc).…”

Automatic Annotation of Change Detection Images

“…Our automatic annotation approach for change image pairs uses a visual semantic embedding model to match the images with the relevant keywords extracted from the publication corpus. When the semantic information about the classes present in EO images is inconsistent or lacking, different solutions have been proposed to extract that information from other sources such as ontologies [ 4 ] or geo-referenced Wikipedia articles [ 21 ]. Another solution is to use visual semantic embeddings by representing the images and text in the space vector space and learning the classes of the unlabeled images based on the similarity between the vector representations across the image and text modalities [ 14 ].…”

“…By integrating an ontology to the segmentation process of pre- and post-disaster images, authors in [ 4 ] showed that overall accuracy went from 67.9% to 89.4% for images of their test area. With a reduced number of samples (200), authors in [ 21 ] demonstrated that using Wikipedia annotations for the task of semantic segmentation, the Intersection-over-Union (IoU) score was 51.70% compared to 50.75% when pre-training on ImageNet.…”

“…This is particularly useful for the monitoring of large or very remote areas where on-site data acquisition is impractical. Indeed, the impact of environmental events such as deforestation [ 1 , 2 ], wildfires [ 3 ], and other natural disasters [ 4 , 5 ] can be assessed with data from EO satellites. With change detection techniques, the various changes that are happening on the Earth’s surface can be automatically detected by analyzing images of a given area taken at different times [ 6 ].…”

“…With change detection techniques, the various changes that are happening on the Earth’s surface can be automatically detected by analyzing images of a given area taken at different times [ 6 ]. Such techniques have been used to monitor loss and disturbances in forests [ 2 , 7 ], to track change in urban areas [ 8 ], and also to map out areas affected by natural disasters [ 4 , 5 ]. When a change detection method neither provides nor needs additional semantic information beyond the change/no-change pixels, it is called a binary change detection method.…”