Hello everybody,
After I did a classification map, there are existed a lot of scattered pixel (class) which no related with neighbor pixels (class). I tried to resolve this problem with Boundary Clean tool in ArcGIS (ArcToolbox--> Spatial Analyst Tools --> Generalization --> Boundary Clean). However, I think that the result is not really good, some wrong as well. Please read it on the "Before boundary clean 1" and "After Boundary clean 1" photos that I enclosed in this question. How could I clear up this case?
Thank you in advandce!
Nguyen
The issue of "salt and pepper" classification results is most frequently dealt with post-classification using a majority filter of some bandwidth (3x3 or 5x5, perhaps). This is probably fine if you are happy with the general classification result and are truly eliminating random noise. This can probably be accomplished easiest in Erdas. Keep in mind that this may be sweeping under the rug, real deficiencies with the classification process.
Envi--Majority/Minority Analysis
https://www.digitalglobe.com/sites/default/files/Multispectral_Analysis_using_ENVI_4_8_and_WorldView-2_Imagery.pdf (last page)
We just looked at this also. If you are using ArcMap, check the Spatial analyst help files for the series of tools to use for the multi-step process. The steps start with the majority filter tool, which is the smoothing algorithm suggested above. Subsequent steps involve the boundary clean and region group tools. Other packages perform similar processes (Erdas, Envi, etc.), as this is a common issue.
Hi Nguyen, if indeed the single pixels aren't meaningful and you want to replace them with their nearest majority neighbors, I'd do the process below. Looks like you're working with classified (i.e., nominal or ordinal) data, and not ratio or interval continuous data, so a low-pass filter is not technically what would be appropriate here.
1) Use the region group tool (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z00000039000000) to identify individual raster regions. This gives you regions that each have a unique integer code, so they're technically zones as well now.
2) The zonal geometry tool (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z000000w4000000) will give you the area of each zone. Single pixels will have area of (cell size) x (cell size).
3) Reclassify the raster, so that those cells that are equal to (cell size) x (cell size) are 1, and all others are NoData (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z000000sr000000)
4) Use the nibble tool (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z00000038000000), with the previous raster as a mask, to classify those single pixels from the original input raster to whatever they're nearest neighbors are. Another option here is the Euclidean allocation tool (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z0000001m000000), after you've wiped the single pixels to NoData using the set null tool (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z00000007000000). Focal statistics, using the "MAJORITY" parameter, is also a way to go (http://resources.arcgis.com/en/help/main/10.1/index.html#//009z000000qs000000).
Hope that helps! :)
Other options could be done by ArcGIS. Converting classified image to vector file (*.shp). It is for you to calculate area of each land-cover type and edit a final map.
Cheers
you might also try a contextual approach, it should help you get rid off this scatterness
Read this one, for exemple
http://docs.lib.purdue.edu/cgi/viewcontent.cgi?article=1438&context=lars_symp
Use Fuzzy convolve option under supervised drop down menu in ERDAS. It deconflicts the classes and generates final classified decision for each pixel
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing