I am actuallly working on a assessment of land use/land cover change using remote sensing in semi-arid context. I have a series of Landsat 5/Landsat 7/Landsat 8 satellite images spanning over 1985 to 2017 period. I have already selected some training sites after a field visit made in early 2017, and I defined enclosed polygons representing thematic land units I am considering for mapping. I ran a supervised classification algorithm (that is, Mahalanobis distance, which performed better in this case than Maximum Likelihood) and I got a thematic map, which have been successfully validated later on using a confusion matrix (kappa was almost 0.84).
I am looking into mapping the same land unit over the previous years. I am very unsure on using the same training sites I already used for 2017 year classification, as I cannot say for sure that those training site were (in the past) in the same state they are nowadays (2017). I was also thinking of derivating a specific spectral signature for every identied land unit I was mapping, then use those signatures to train a classifier on the 2017 image and then let it run on older Landsat images. Any thoughts ?
How can I accurately perform this task ? I am actually familiar with ENVI software (v5.1). Reference to peer-reviewed articles would be fine.
Regards,
Roland.
Hey Roland,
Here are few suggestions which may help you
1. Make sure that all images are normalized before classfication.
2. You may try Random forest / support vector machine (available in ENVI). The machine learning classfier works better.
3. you may try PCA on Landsat Band and check classfication on PCA 1,2 and 3.
4. If you are looking at change detection then you may want to consider gain and loss pixels to avoide misclassfication.
- Badges
- Science topic
- Similar topics
- Geoscience
- Remote Sensing