We are trying to classify landsat 8 surface reflectance product for landu usse/ land cover change detection. Conventional classification methods didn't worked well and could not give desired accuracy because the features on ground are very mixed with each other and also bands are significantly correlated to each other. To reduce this correlation between bands, Principal Component Analysis is used. But we are unable to incorporate it with other information in decision tree classifier.
Hi karki
First check for these aspects
1. Are all your images have the same spatial references and resolution
2. Are all your images have the same size in term of number of lines and rows. Depending on the software you are using, PCA images can have lower sampling size than original images.
If it's not the problem, look for the software documentation to verify if your data meet all the requirements to run decision tree
Best regards
Consider looking into mutual information/(gaussian) mixture models instead. PCA likely will be no good for your task. Still, on the technical side of things you can try running your usual metric classifiers/algorithms with Mahalanobis distance instead of L2, that'd be the same as the L2 on PCA'd data.
I'd still recommend steering away from context-dependent processing steps if you want your algorithm to generalize well, which you should with temporal series.
Dear Sijan,
Which software do you use? If R, you can use package 'raster' for raster data operations and package 'vegan' for PCA. You can simply calculate the new raster layers corresponding to the PC axes, and then do the cluster analysis using these new raster layers.
HTH,
Ákos
You may need to consult the works of Late Dr. Trevor Kletz for the ground concentration model of air pollutants in terms of concentration and distance coordinates for the various effluents. Inherently safer design is another interesting work of the professor that can be of assistance.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training