The concept of transferability and extendibility in object rule-set based image analysis

摘要

The purpose of this study was to test the performance of different object-based classification models, using different data sets, for urban land cover classification at large mapping scales. The study area was the city of Cambridge, UK. The acquired data was a true-colour aerial photography (0.125 m RGB) and VHR QuickBird satellite imagery. Initially, a rule-set classification model was developed, in three sample areas, using the aerial photography and eCognition software. The performance of the classification model was statistically evaluated with the error matrix approach and API as reference data. The mean overall thematic accuracy was 91.3% when simple binary thematic maps of vegetated vs. non-vegetated surfaces were compared. The model was also applied on satellite data of the same areas. The overall thematic accuracy was 79.2%. The 12% decrease in thematic accuracy could indicate the necessity of also using the NIR band. A new object-based classification model was developed based on the additional spectral information of the satellite data, i.e. the NIR band and the NDVI that are useful to enhance sensitivity for detecting vegetation. The statistical analysis of the model indicated a mean overall thematic accuracy of 80%. The RGB object-based classification model showed the potential of transferability as it has been successfully applied on different data sets although there was a significant decrease on the performance. However, the development of a new model using the spectral information of the satellite data did not improve the overall thematic accuracy indicating the limitations of the spatial resolution of satellite data.