Land cover classification Change detection

4.2 Land cover classification

The actual land cover classification result for the study area is shown in Figure 5. More detailed figures of the results can be found in Ahokas et al. 2016.The segmentation and selection of training segments were carried out as described in Section 4.1. The Random Trees implementation available in eCognition was used to create 1000 trees for classifying high segments into buildings and trees and 1000 trees for classifying low segments into asphalt , gravel , rocky areas and low vegetation . In the classification of high segments, both DSM and intensity features were used. The classification of low segments was based on the intensity features. As a postprocessing step, buildings smaller than 20 m 2 were reclassified as tree . Most of such very small buildings are misclassifications. Figure 5. Result of land cover classification class High object includes a few very small segments that remained unclassified in further building tree classification.

4.3 Change detection

The change detection was demonstrated in a smaller area. The results are presented in Figure 6. The rules and threshold values applied in the change detection are listed in the following. DSM_DIF is the absolute value of the difference between the mean heights in the new and old DSMs; h is the height of the segment, i.e., the difference between the mean heights in the new DSM and DTM; Ratio Ch2 is the mean intensity value in Channel 2 divided by the sum of the mean values in all channels. The feature Ratio Ch2 and its threshold value were obtained from the classification tree analysis. Levels refer to segmentation levels. The main steps of the change detection included:  Analysis of building segment level 1 derived from the old map segments correspond to buildings: o DSM_DIF ≤ 2.5 m - Old building OK o DSM_ DIF 2.5 m and h ≤ 2.5 m - Old building demolished o Otherwise - Old building changed  Segmentation of the new DSM outside the old buildings and analysis of this DSM segment level in change detection we used the minimum DSM because it is advantageous in the analysis of buildings. o h 2.5 m and Ratio Ch2 0.3005144 - New building  Combination of all buildings into building segment level 2 and analysis of this level. Building segments connected to each other were merged to one segment. o Area 20 m 2 - unclassified to remove small erroneous building segments o Relative area of sub-objects Old building demolished ≥ 0.1 - Old building demolished o Relative area of sub-objects Old building changed ≥ 0.1 - Old building changed o Relative area of sub-objects New building ≥ 0.9 - New building o Relative area of sub-objects New building ≥ 0.5 - Old building changed o Otherwise - Old building OK Finally, the original shape of buildings classified as Old building OK was retained by converting them back to sub- objects and removing building parts classified as New building . Figure 6. Automated change detection for buildings. The figure shows the old DSM and old building vectors left, the new DSM middle, and the change detection result right. The legend applies to the change detection result. Some areas were excluded from the change detection analysis due to missing data in the old DSM. Original building vectors © the City of Espoo edited by the FGI. This contribution has been peer-reviewed. doi:10.5194isprsarchives-XLI-B3-323-2016 328

5. DISCUSSION