2

Is it possible to do accuracy assessment using polygons as reference data as against point data?

I have field-identified polygons as data for training and reference (from a forest-agri mosaic) that have been identified from field. The regular post-classification accuracy assessment method is to use random points and check background image. But instead can I use my field polygons as reference data, instead of points. Problem being, I cannot correctly interpret the land use on the background image or even on googleearth since it is complex forest-agri mosaic. All I have is my reference data which are polygons.

I could convert each polygon into a point and do it, but that defeats the purpose, I feel. the idea is to try to get the software to recognize all pixels in every reference polygon as a single class. Can this be done? If yes, how? I am using ERDAS 2014, but I also notice the QGIS 2.6 plugin for accuracy assessment.

2

It is better to select a reference that is consistent wit the one used to create the map (points for pixel-based classification, polygons for object-based classification), otherwise it is more difficult (or even impossible) to unambiguously assign a label to the validation entity.

In your case, there is clearly a gap between the classification and the validation. But if you need a comparison it should be done at th polygon level. Therefore you should translate your legend into aggregation rules (with a LCCS type legend this should be possible) and use it with some zonal statistics at the object level. Once each polygon has a label that correspond with the pixels that it contains, you can get a grip of the quality of your pixel-based classification (beware that this would not correspond to the overall accuracy of your data, because the polygons DO NOT exist in your map at this point, however it should be useful to document what you've done)

1

If I understand you correctly, you have polygon data for training and validation. Aside from common methodological technicalities in using point data I see no reason that you cannot use your polygon data. This conceptually treating your polygons as a discrete spatial process with an expected spatial homogeneity in the validation. There will most certainly be some uncertainty issues so, the error will not be a very precise measure and I would take it with a grain of salt.

A simple approach would be to convert your polygon data to a raster, with the raster values matching the same domain used in the classification. Also, make sure that the snap (origin coordinates) and extents match (row, columns, coordinate extent). The background, areas outside the polygons, should be set to NoData.

A nice, and free, software for this type of validation is the Map Comparison Kit. A common statistic, that is conveniently implemented in MCK, is the Kappa, which is a chance corrected percent correctly classified. The authors of the MCK graciously compiled several relevant papers on categorical comparison and validation methods implemented in the software, into a single pdf. I would highly recommend reading through the papers.

1

Yes, it is possible. You can use the intersect tool in ArcMap (under analysis tools - overlay). Add your point and polygon into input features. Use point as the output type. And you can do your accuracy assessment using the table of the resulted intersected map.

Note: Make sure that you can still recognize the attributes of the point and the POLYGON within the intersected map for your accuracy assessment purpose.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.