Semi-Automatic Object-Based Building Change Detection in Suburban Areas from Quickbird Imagery Using the ERDAS Imagine Objective Software

This paper aims at effectively detecting the building changes in bi-temporal very high resolution satellite images. Many applications require detecting structural changes in a scene over a period of time. Change detection of man-made objects using remote sensing images has many applications such as city planning, informal building detection and disaster management. For this paper, two bi-temporal multispectral images from the QuickBird satellite are used, depicting a region of south-eastern Attica. The implementation of the semi-automatic change detection of the two images is made in the object-oriented environment of the ERDAS Imagine Objective. The method is based on the comparison of two independent classifications. The independent building extractions are the result of tree-processes in the Objective. The first step for each feature extraction in Objective involves the system training by defining background and non-background training samples. Afterwards, Objective creates a probability layer which presents the single probability of each pixel for being a building, based on the training samples. Then, the creation of objects is followed by a raster object operator such as segmentation. Subsequently, the created objects are processed by applying a variety of functions including probability, size or morphological filters. This is the last raster level since the next level converts the objects from raster to vector form. Consequently, the objects are processed by operators which reshape the existing objects, eliminate these who do not meet certain criteria, combine multiple objects to a single or split object into multiple new vector objects. The next level is to perform classification on the vector objects. Vector object classification involves specifying one or more cues which are used by the Object Classifier. Cues include metrics which certain properties of vector objects are measured. The Object Classifier uses the cues to assign a probability to each object in a group of vector objects. Finally, the last level includes operators which typically clean up the set of vector objects to produce a nice final output. Some Vector Cleanup Operators use the probability attribute generated by the Vector Object Processor in the operation. The implementation of the change detection is carried out by the comparison of independent building extractions using an operator that computed the probability of change for every object, taking into consideration the probabilities of the objects for both dates, and the distance of centroids of the polygons between the two dates. The results of extraction are very satisfactory since the correctness is 85.9% and 85.2% and the quality 67.0% and 67.5% for the first and second dataset respectively. The change detection do not produce proportionally successful results since the correctness is 22.3% and the quality 13.4%. The reasons which contribute to the lesser accuracy of the change detection results are primary, due to the mediocre geometric registration of two images, the limited spectral information of the data as well as the large extent and complexity of the study area.