I want to identify new houses built from 2015 to 2016 using satellite imagery (MSI and SAR)
Would anyone have any suggestions how to go about this?
I was thinking if I did a supervised pixel classification of both images (2015 and 2016) and used classes such as vegetation. soil, water, forestry, built up to create 2 landcover classifications. If i think converted these raster classifications to vector data sets, and exported only the ‘built up’ polygon from 2015 and 2016, and then clipped the old 2015 ‘built up’ vector set, to the new 2016 ‘built up’ vector set, the ‘new built up areas’ would be left over.
your approach would be a so called post-classification analysis. If requires a bit of work before you can extract the results but if the classifications are reliable (calibrate your data to make them comparable) the results should be the most correct.
For a quicker estimation of changes you can use
one band of a multispectral sensor
one band of a SAR scene
(make sure they are from the same time of the year to minimize phenologic variation)
and compare the changes in reflectance/backscatter intensity. Especially for the SAR image, new houses shoud result in significantly higher backscatter.
You can even use the change detection module of snap for that.
For SAR the single-, double- or triple-bounce reflections from a building are usually very bright, so I don’t think the phenologic variation matters that much. The issue here is though that the reflections from buildings are highly directional, which means that suitably oriented houses will not be visible at all. This issue can be mitigated by using both ascending- and descending scenes in the analysis.
In addition to analysing intensity, looking at a long temporal-baseline (say 6 months or more) coherence images is worth trying (bright permanent structures stay coherent for a long time).
In SNAP you can to calculate log ratio and segmentation bands… Go RADAR— SAR Applications — Change Detection — There you obtain the Log Ratio Band, after with Band Maths you will make the segmentation using the mean and sigma obtained from the Log Ratio Band