Land cover classification and Change detection technique

[Daniela_06]

Hello,

I need to do a land cover classification of satellite images from diff years 86,91,98,02,12 from different providers like ASTER and LANDSAT and I have several questions regarding some issues I am presenting. Hope someone can help me.

1. Do the Landsat images need to be ortho rectified, they are L1T images TM and ETM+ if so, can I use the ASTER DEM of 2011 that I have to ortho rectify all of them? no matter the year? An if the DEM is in a different resolution that the 30m how should I do this?

2. I need to do some atmospheric correction and enhancement of the bands that i need to use that will be the false colour combination to be able to distinguish vegetation. This enhancement should help me with the classification procedure right?

3.For the year of 1992 I have a shapefile with different polygons of the different classes. I been meaning to use this polygons as ROI in ENVI but I cant seem to make the classification work. (Supervised classification Maximum likelihood) . When loading the shapefile as ROI or EVF and using it as ROI's for classifications is returning a white screen. How can I fix this?

4. A 1992 is the only year I actually have some a priori info I want to (by visual analysis) try to get some signatures or ROI for the other year to do the classifications, however, in some cases land covers or pixels that represent a land cover are very similar to pixels that represent other land cover and this is producing a horrible and totally inaccurate map. What level of treshold should I set to that one region does not get ¨confused¨by another region?.

Please I hope someone can help me, right now I am trying to classify 14 classes. in the case the maps are really complicated then I may focus in reducing the classes and maybe concentrate in levels of vegetation, but I would like to give it a try.

The area of Study is a semi-arid area where all the land covers are similar. I know that the shapefile that I have is key to this and would like to make the classification work. The attribute table of this shape file is like this.

ID/CODE/POLYGONS where the code is the one showing the type of land. Please help!

[rahmansunbeam]

1. Do the Landsat images need to be ortho rectified, they are L1T images TM and ETM+ if so, can I use the ASTER DEM of 2011 that I have to ortho rectify all of them? no matter the year? An if the DEM is in a different resolution that the 30m how should I do this?

If the subjected area has significant amount of terrain and geometric distortion then you'll surely need orthorectification. If it is ENVI you are using, try this

http://invisibleroads.com/tutorials/envi-image-orthorectify.html

2. I need to do some atmospheric correction and enhancement of the bands that i need to use that will be the false colour combination to be able to distinguish vegetation. This enhancement should help me with the classification procedure right?

Any enhancement will make the image loose something to gain something. The atmospheric correction will reduce the impact of atmospheric haze locally, but will effect the result globally (unless you limit the working area with some roi). Color composition are only for visualizing various features of an image, it doesn't effect any operation. There are many ways you can delineate vegetation from Landsat, search this site for them.

btw, what is it about 'all the land covers are similar' ? I'll suggest to use as little enhancement as you can.

3.For the year of 1992 I have a shapefile with different polygons of the different classes. I been meaning to use this polygons as ROI in ENVI but I cant seem to make the classification work. (Supervised classification Maximum likelihood) . When loading the shapefile as ROI or EVF and using it as ROI's for classifications is returning a white screen. How can I fix this?

There could be many prob, maybe the projection do not match, or the shp 's are not good enough. Draw ROIs within ENVI then try. (just guessing !)

4. A 1992 is the only year I actually have some a priori info I want to (by visual analysis) try to get some signatures or ROI for the other year to do the classifications, however, in some cases land covers or pixels that represent a land cover are very similar to pixels that represent other land cover and this is producing a horrible and totally inaccurate map. What level of treshold should I set to that one region does not get ¨confused¨by another region?.

Please I hope someone can help me, right now I am trying to classify 14 classes. in the case the maps are really complicated then I may focus in reducing the classes and maybe concentrate in levels of vegetation, but I would like to give it a try.

A common problem in pixel based classification system. Mixed classes/pixels are always hard to identify. You can,

-combine both super and unsupervised classification for each class. Common practices is to collect signatures from unsupervised and use it for supervised classification.

-play with settings. Classes do not follow same setting.

-draw and collect class signatures carefully. That's more like on-screen digitization - very boring. If it's Erdas, i'll suggest collecting seeds, don't know anything similar in ENVI.

-sub-pixel classification, try it

-give object-based classification a try. You'll have more accuracy in less time.

[Daniela_06]

Thank you very much.

So? Orthorectification can be done using ANY year DEM over ANY year image? I dont think my images need Orthorectification, but no harm can be done if i do it anyway. right?

Regarding the classification, im going to follow your advice and see what happens.

[rahmansunbeam]

So? Orthorectification can be done using ANY year DEM over ANY year image? I dont think my images need Orthorectification, but no harm can be done if i do it anyway. right?

the only info is operational is the height, which use to be a stable feature. But using updated products are always recommended.

[Daniela_06]

Hello, I figured out my classification procedure and is GREAT! I was able to work with the shape file i had for the 2002 year and basically by visual analysis I will have to do the classification of the years that i dont have apriori information whatsoever.

I have another question:

I am collecting the signatures, or training samples myself, by comparing by visual analysis (looking at the map) of the 2002 classification I have. However, can you explain me the procedure i need to follow to assess the accuracy of my samples and the accuracy of my classification?.. vs what am I comparing or creating a matrix if I dont have ground control points?

Accuracy in base according to what and what do I need to report? I am using ENVI or ERDAS software with TM, ETM+ and ASTER IMAGES. THANKS!