sorry for late, Wei Hua (I did not see yours until today),
Last 7 years, we firmly moved from the classic image analysis workflow to object-based, and have been gradually using Erdas Objective to update GIS database and (landuse) change detection over time in operation, meanwhile closely watching ESRI and other partners on this, in particular, seamlessly with imagery within Mosaic Dataset or from Image Analysis:
- For ArcGIS Pro 1.4 (10.5), SVM offers better landuse classification results, but still do not directly produce feature extraction (roads, buildings, etc.), because of no 'effective' generalization tools available in ArcGIS (refer to the Doc).
- For Erdas Obective, it uses 7-layer vision framework (especially with Bayesian network models), which works very accurately and effectively on the extraction and cleanup of features (roads, buildings,...), especially if only interested in new features and changes (which are real scenario in GEO Intelligence). Please refer to their Doc. Worth to investigate if hierarchical kernel descriptors framework with SVM (already available in 10.5 & 1.4) can produce better results. It looks true, which is based on many researchers.
In addition, for the Segmentation & Classification toolset from 32-bit ArcGIS 10.4/10.5, the performance is very poor, which looks that it does not really support the multicore algorithms. Personally, it should be investigated to ensure functional on both multicore CPUs (32-bit, 64-bit) and even GPUs (64-bit) as well....
Regards,