Urban Dynamic Monitor

Post-classification adjustment

The change detection products provided by UD processor are particularly suitable for statistical description of urban expansion. The areas with dynamic development within the cities have been identified. In case detailed monitoring/mapping of the urban areas is required (currently usually based on semi-automated pixel-based or object-based change estimation and visual editing) the post-classification manual adjustment of the change detection products is needed.

The main aim of the post-classification editing is to demonstrate the possibilities of further use of UD processor products. Also this adjustment leads to the increase of the overall accuracy. On the basis of two criteria (dynamic expansion of urban area and availability of reference data) the region of size 6 × 6 km within the eastern part of Surabaya AOI was selected. The final product from the period of 2009-2014 was manually adjusted and evaluated using the Google Earth VHR imagery.

Figure 13: UD processor result (left) and result of manual editing (right)

The effect of the manual post-classification editing was assessed by the same validation method used for UD processor results evaluation.

Layer Number of samples Overall accuracy (%)  Error of commission (%)
Before manual editing 100 58 42
After manual editing 100 80 20

Table 4: Overall accuracy for selected subset in Surabaya AOI 2009-2014

Results of accuracy assessment presented in table above prove that the manual editing improves the change detection results significantly. Accuracy of the UD processor result after manual adjustment increased to 80 % and error of commission has been reduced by more than 20 %. The main errors that were corrected during manual editing were overestimation of changes – the change was detected in place where in fact no change occurred (isolated small polygons) or extent of the change was larger than in reality (incorrect delineation of change). The isolated change polygons were located especially in the existing urban areas or agricultural land. On the other hand the wrong delineation was identified on the edge of expanding industrial units. Limits of spatial resolution and interpretability of Landsat satellite data are there evident.

The main benefit of the use of UD processor product for manual adjustment is the reduction of manually inspected area. The sample area that has been subject of this test had a size of 6 × 6 km. Thanks to the use of UD processor products the area to be inspected fell to the 10 % of its original size. Likewise the time required for this processing also decreased. The total time needed for manual adjustment depends in particular on the landscape variability, number and quality of the input data, but also on experience and knowledge of the interpreter.