How do I classify a highly heterogeneous built-up area?

Question

I am using Landsat 8 imagery for supervised classification. I have used maximum likelihood classifier to classify separate classes such as forest, grassland, agriculture, plantation, water body and built-up area. The built-up areas in my study site are small towns with a lot of green cover, and therefore, had been classified as either one of the vegetation classes stated above.

Should I create a polygon and clip those areas where these towns are located and just classify them as built-up area? Also, please note that I will be doing change detection on this classified image and don't know whether it would be advisable to use this method to 'manually' classify these built-up areas.

Answer 1

That would likely be your best bet, outside of object-based classification.

However, another option would be to use a more advanced algorithm for classification, such as the Support Vector Machine, or a decision tree-based model, as these will allow for more detailed classification.

If you are forced to use manual classification, be sure to be consistent in your classification, or your results will be completely moot. This article has a short discussion on this problem: doi: 10.1109/TGRS.2014.2321423.

If your software cannot perform more advanced classification, download SAGA for free, it has a good selection of algorithms to use.

Answer 2

There are two facets for this issues: one is conceptual and the other is technical.

The conceptual issue is that you need to define what you mean by built up area in your legend. A good starting point is to make the exercise to build a LCCS legend. It is indeed necessary to define what you want as a model before trying to map it. Note that in your case, a thematic map could be suboptimal but you could try to estimate the proportion of impervious surface inside each pixel. Furthermore, you should choose if your landscape is best represented with pixels of polygons (what would be your minimum mapping unit). The "scale" (or granularity) of your "heterogeneity" is therefore very important. Asking yourself if the heterogeneity is below or above the pixel should help you to take the good decision. This decision is particularly important if you want to perform a change detection, because you do not want to compare apples and pears.

From the technical point of view, manual delineation is great as long as you do not have a large area. Otherwise you could use image segmentation (if you are interested in a complex legend). There are segmentation tools available from third parties within QGIS (SAGA, OTB, GRASS). Note that Image segmentation will also aggregate pixels in the other land covers: this will reduce the salt and pepper effect, but you will loose some details. Alternatively, if your urban area is characterised by its texture due to heterogeneity, then you could use some pixel-based texture (e.g. a simple high pass filter in a 3*3 window).

As a remark, as mentioned by others, you should avoid Gaussian maximum likelihood if (i) you have a very large number of features or (i) the spectral signature of your classes id not normally distributed.