Colloquium

Automatic change detection with SPOT 6

Organisator Laboratory of Geo-information Science and Remote Sensing
Datum

di 25 augustus 2015 14:30 tot 15:00

Locatie Lumen, building number 100
Droevendaalsesteeg 3a
100
6708 PB Wageningen
+31 317 481 700
Zaal/kamer 2

By Mauro Garcia Esteban (Spain)

Abstract
Cities are like living organisms, growing and changing continuously, making the mapping of them a continuous task. Efforts have been done before in the field of automatic urban change detection by using different imagery sources as inputs. Usually, aerial photographs are used as inputs, but their high cost makes it hard for small municipalities to acquire them regularly. A cheaper alternative can be found in VHR satellite images, which are continuously increasing in availability. Supported by that, the aim of the study was to analyse the possibilities of detecting changes in an urban environment by using SPOT 6 images as input, being the final goal to develop a method for cheaply updating the new Dutch large scale topographic map, required to every Dutch municipality. This new topographic map should cover the whole area of the municipality with polygons classified as one of the main BGT classes: water, road, house and green areas. Next to those classes, even though it is not required in the new topographic map, the class trees was included in this study, were the trees were point features. The choice for SPOT 6 images for this study responds for them being openly accessible via the Dutch Space Agency.
Having a working and effective methodology for automatic change detection could help small municipalities to update their topographic maps regularly. Even though the developed methodology does not update the map automatically, it indicates which objects have lost their validity and should be checked. For doing so, different methodologies were tested for the area of Drents Dorp, a residential neighbourhood in the municipality of Eindhoven, in the South of the Netherlands. The base methodology consisted of comparing an image from 2012 with an image of 2014 and looking for areas that had changed in class between those years. The areas pointed as changed were compared with a dataset showing the real changes in the BGT objects. The accuracy of the change detection was measured in terms of objects, in terms of classes and in terms of area analysed. Two corrections were added to the base methodology in an attempt to correct for situations that could hinder the accuracy of the change detection. One of the correction was for dealing with temporary changes in the objects, where an extra image was added to the comparison. The other correction dealt with the areas of the objects covered by shadows or occluded by tilted objects. In this last correction, a DSM was used for detecting the areas and masking them out from the analysis. Finally, a combination of the corrections was also implemented. The base methodology for the class trees was similar to the one used for the BGT objects, but since the trees are point objects, they needed to be buffered in order to be compared in the images. Also, a different validation dataset was used for assessing the accuracy of the methodology.
The results for the different methodologies showed that the corrections improved the accuracy of the change detection, being the highest accuracy that of the combination of corrections, reporting a 95.6% of overall accuracy in terms of polygons. In terms of classes, the class reporting the highest accuracy was the class house (97.1% in the combined corrections methodology), while the lowest accuracy was for the class trees (58.4% in the basic methodology). Class road and class green areas obtained 93.5% and 92.2% accuracy respectively for the combined corrections methodology. Class water could not be tested since there was only one object of that class in the area of interest. Despite the high accuracy values obtained for the different methodologies, only 65.5% of the total area is visible due to shadows and occlusion, remaining uncertain the occurrence of changes in the rest. In relation with the accuracy for class trees, the low results seem to be related to the buffering of the points. For some trees this might work, whereas for small trees that are not dominant in the buffered region, this could lead to the analysis of a different class. Next to that, the inputs of the study were not optimal: the BGT needed to be created based on different land use datasets since there is no BGT map yet for that area, the satellite image for 2012 needed to be from Formosat2 due to unavailable SPOT6 images for that time, and the DTM used for calculating the shadows and the occlusion was only available for 2012.
In view of the results, it could be concluded that detecting changes in an urban environment by using SPOT 6 images as input is possible, even with high accuracies. To do so, the elaborated methodologies could be used for finding out the obsolete or out-of-date areas in the BGT maps, speeding up the updating process. Nevertheless, improvements to the methodologies are needed, especially for the classes with low accuracies, as well as validation of the methodology for different areas.