I am trying to make a supervised classification on a Landsat 7 image using the QGIS 3.16.1 version. I am using the SCP plugin to do then classification. I am doing the classification using the RGB band 4-3-2 and the maximum likelihood classification.The problem I have is that in the final product the bare or dry fields appear like urban areas. Does anyone know how to solve this?
-
1This is a very common problem. Why not use the additional bands available with Landsat data?– Aaron ♦Dec 11, 2020 at 21:40
-
So I should select the sample for urban areas with the 4-3-2 band and agricultural areas with 6-5-2 maybe ? Do you have something else to suggest me or is there any other way to fix it?– kopa23Dec 11, 2020 at 22:33
-
1Use all of the available bands for your classification. You may also want to consider incorporating indices such as NDVI, NDBI, and NDWI.– Aaron ♦Dec 12, 2020 at 1:40
1 Answer
There are several issues you need to address to improve the accuracy of your land cover classification.
Make sure you have a large enough sample of training data and that your training data is distributed equally among all of your classes. A good rule of thumb is to use at least 50 samples for each class for areas <1 million acres and fewer than 12 classes (Congalton, 1988)
Use all of the available bands for your particular sensor. If you are using Landsat 7, certainly incorporate bands 1-5 and 7.
Also include indices such as NDVI, EVI, NDBI, and NDWI. More details here. These will help to further differentiate classes.
Consider incorporating a non-parametric decision tree classifier such as Random Forests.
Reference: