Goodmorning,
I’m trying to do a land cover classification of SLC Sentinel-1 dual-pol data. I processed the product with the following steps:
apply orbit file
calibration
deburst
TC
Then I produced GLCM and I applied a speckle filter to the Terrain-Corrected product, in order to make an H-alpha decomposition.
After that i tried to include the three products (GLCM, H-alpha, speckled data) into classification data but, both with KNN and RF algorithm the image I obtain isn’t all classified, as in the screenshot which is following:
I do not know what is going on with your error, but speckle filtering, GLCM and H-alpha etc. should be performed before terrain correction. TC resamples the data and changes the original statistics leading suboptimal/wrong results.
no, the classifier just does not find a class with a high confidence.
You can right-click on the results raster and lower the threshold in the valid pixel expression (or remove the expression completely) to have a class assigned to all pixels regardless of the confidence.
I used the bands produced by the pre processing of the image, the bands produced by GLCM (all the available bands), and the 3 bands from H-Alpha decomposition. More than 40 features (I think they were too much).
Now I only included the bands produced by the pre processing. I’ll try by doing a PCA on GLCM and add PCA to the features of my classification.
Is this workflow right? I’m new to this and I’m not always sure of what I’m doing
To be honest, I would have expected the Random Forest algorithm to “sort out” the unnecessary features (or at least deal with the redundancy inherent in the input data).
How many trees did you use? If you use more than 40 features, processing at least 100 trees would make sense to fully exploit the principle of permutation of the features.
But as a work-around, reducing the number of input rasters by applying a PCA is totally fine.
I was using the standard setting (10 trees), but now I tried with 200. Passing the speckled image, the GLCM bands and the H-Alpha bands to the classifier, I still have the same problem of unclassified pixels.
I made a lot of tries passing different kinds of raster to the classifier. The problem of NaN pixels appears only whenever I give glcm bands to the classifier (even only one). My question is: is there something I have to know (such as some settings I have to modify) about glcm generation? I kept the default settings in the box.
that is what I suspect as well. Some textures derived by the GLCM module produce NaN pixels. You can have a look if they coincide with the pixels which were not classified.
If so, you have to uncheck the option in the band properties which assigns 0 as no-data and save the product. The corresponding pixels are then black.
I have been naive and I didn’t see that the non-classified zones are NaN in GLCM. So there’s some option in the GLCM box which masks out these zone (mainly seas). could someone explain which is this option?
If I remember correctly, they are not NaN but simply 0 (which is defined as nodata in the properties). If you unckeck the use of nodata and save the product again, the classifier might handle it.
