Say that you have Sentinel-2 data and want to derive land cover classes from that data. You adopt a common pixel-based classification approach, say, Random Forest. You design a suitable classification scheme and give the training samples to the algorithm. The classification result is straightforward: all the classes are identified in one go, based on the algorithm’s decision for each pixel.
The output will be sound in distinguishing among classes such as water and vegetation as their spectral signature is way different and can be easily differentiated. However challenge lies in determining between categories such as grassland and cropland, bare soil and built-up land.
To this effect, recently, researchers have proved that adopting the hierarchical classification approach can prove more beneficial rather than the standard approach.
What is meant by the hierarchical classification approach?
In the hierarchical classification, first, the data is segregated into more easily separable classes, and then in further attempts, the challenging courses are identified. Here the classification process is run iteratively. Finally, all the individual outputs are combined into one resultant image.
Any limitations?
Yes, it is to be noted that the hierarchical classification approach increases the complexity of the classification process and thus consumes more time.
Reference:
Waśniewski, Adam, Agata Hościło, and Milena Chmielewska. 2022. “Can a Hierarchical Classification of Sentinel-2 Data Improve Land Cover Mapping?” Remote Sensing 14, no. 4: 989. https://doi.org/10.3390/rs14040989
