I have a Landsat time series from which I derived NBR using bands 4 and 7, and delta NBR (to show changes). I then set a threshold based on the mean and standard deviation delta NBR values to reclassify rasters to change (1) and no change (0). The result is a 28 series of images representing change events between years for a study area of about 5,000 km^2. This was all completed in R, and I have the output in raster and vector (polygons) formats. Only pixels labelled as change (1) were used to generate objects (vector polygons).
The output is reasonable for many of the change types occurring in my study area - but I have encountered an issue with roads and other linear features. Instead of the ideal continuous linear (or curvilinear) feature, my output includes many discontinuous features with gaps (as shown in the two examples in the image below).
I have tried to use the Radon transform (PET package in R), but have not been successful. I am looking for a gap filling solution that can be implemented in either ArcGIS or R, using either raster or vector. If the solution is also good for noise removal, all the better! Next steps after gap filling are noise reduction, then classification using Random Forests.
(BTW, this is for an undergraduate project so my level of expertise is moderate)