Automatic multi-scalar interpretation of multi-temporal remote sensing data (2011)
|Duration:||4/2008 - 3/2011|
The goal of this project is to develop a method for automatic interpretation of multi-temporal remote sensing data of different sensors. It deals with optical satellite images, which might have different scale. The motivation is to link the interpretation results of expensive high-resolution images with the interpretation results of cheaper low-resolution images. For this a strategy has to be developed, that is able to combine images of different resolutions in an arbitrary way.
Approaches for the interpretation of multitemporal data aim at detecting the occurrence and the type of changes. Most of the existing methods are restricted to the comparison of monotemporal classification results, temporal context knowledge is considered very seldom.
The use of spatial and temporal context knowledge might enhance the classification quality. One popular method in the field of pattern recognition is Markov Random Fields. A label is assigned to each primitive based on its features. These labels interact with their neighbourhood to determine the most probable configuration of all primitives. Conditional Random Fields (CRF) differ in a way, that not only the labels, but also the features of neighbouring primitives interact.
The gain, that can be achieved by considering context knowledge, can be seen in figure 1. The classification results for four classes (settlement, industry, forest and agricultural) of a RapidEye-scene, that was analysed together with two Ikonos-scenes, are displayed.
Figure 1: a) reference, settlement highlighted in red, industry highlighted in blue, forest highlighted in green, agricultural without highlighting; b) result of a ML-classification – no consideration of context (64,2% correct segments); c) result of a monotemporal CRF-classification – consideration of spatial context (74,5% correct segments); d) result of a multitemporal CRF-classification - consideration of spatial and temporal context (79,1% correct segments)
Current work is focused on an expansion of the approach to multi-scale data. For that purpose, knowledge of the appearance of the different land use classes in a variation of scale needs to be integrated. This encloses information, if a class is visible in different scales and how it can be automatically extracted in the respective scales. With the multi-scalar modelling it should be possible to make predictions for consecutive interpretations. Moreover conclusions can be drawn, which objects, that are only visible in high-resolution images, exist in images of lower resolution. With the combination of the interpretations of the diverse images a significant improvement of the overall interpretation results is expected.
Hoberg, T.; Rottensteiner, F.; 2010 (2010): Classification of settlement areas in remote sensing imagery using Conditional Random Fields: IntArchPhRS vol. XXXVIII, part 7A, Vienna, 2010, pp. 53-58 | file |
Hoberg, T.; Rottensteiner, F.; Heipke, C. (2010): Classification of Multitemporal Remote Sensing Data Using Conditional Random Fields: 6. IAPR TC 7 Workshop on Pattern Recognition in Remote Sensing, Istanbul, 2010, 4 S. | file |