Remove raster connected regions below a threshold
Connected component analysis
Files processed in browser
Effectively removes noise and small regions
Upload the raster file to process
Set threshold and connectedness
Regions smaller than this will be removed
The Sieve Filter removes speckle from classified rasters: it labels every connected region of same-valued pixels, counts each region's size, and deletes regions smaller than the threshold you set, measured in pixels. You decide whether pixels connect through 4 neighbours (shared edges) or 8 (edges plus corners), and the connected-component analysis runs entirely in your browser.
It is the cleanup pass nearly every classification workflow needs - a satellite land-cover map scattered with one-pixel water spots, a noisy crop map, or a scanned soil map all turn into presentable products once the small fragments have been sieved away.
Upload the classified raster in any of the supported GDAL-readable formats.
Set the threshold - connected regions with fewer pixels than this are removed, so 2 is the gentlest possible pass - and choose 4-connected or 8-connected neighbourhoods; 8 also joins pixels that touch only diagonally.
Run the filter; the tool reports how many small regions were removed, then download the cleaned raster.
AThey are set to the raster's nodata value rather than merged into a neighbouring class. If you would rather have those holes filled with surrounding class values, follow up with the Fill NoData tool afterwards.
AStart just above the size of the noise you want gone: a threshold of 5 removes regions of up to 4 pixels, while 50 eliminates specks of about 7 x 7 pixels. Increase gradually - an oversized threshold starts eating legitimate small features such as narrow ponds or single buildings.
AWith 4-connectedness, two diagonally touching pixels belong to different regions; with 8 they merge. Eight connections dissolve staircase-shaped fragments that only meet at corners, so the same threshold usually removes noticeably more area.