landsat 5 and 8 classification problems

[styfler]

Hi guys, hows doing?

 

A old problem that I have with both landsat 5 (bands 543) and landsat8 (654) in classification of urban areas is a confusion of the classification (maximum likehood classifcation) about some low albedo features, mainly shallow water and asphalt. I tryed a thousand times, with the most accurate samples I could do (fusion image with pancromatich), both at envi and arcmap.

 

The only software I could obtain a better result was Ecognation with segmentation, but I only have the trial version, so I can't export the file.

 

How can I minimize this problem to have a better classification at least of this features: urban, soil, grass, forest and water. The classifications are failing at the edges shallow areas of some huge urban lakes that I'm trying to study.

 

Thanks

[Arhanghelul]

in ENVI and ERDAS you can use a sub-pixel endmember classifier (ex: SAM).

Yes, is difficult, probably impossible to separate some classes in urban areas based only on spectral values.

An OBIA approach, using eCognition can produce better classification results compared to traditional per-pixel methods (ex: maximum likehood and ISODATA). This is because eCognition creates object segments, and you can use the shape, texture or position of the objects in you satellite image for superior classification results.

 

Conclusion: first, try to use a sub-pixel method, with a neural network method, and if the results are not satisfactory, try to use eCognition.

You can find the eCognition software here, on this forum, with the crack/medicine, Give a search on the forum.

Edited June 12, 2014 by Arhanghelul

[deepgis]

You can use NIR Band values to separate shallow water bodies from urban areas. It would be better using ancillary data layers (such as elevation, slope, roads etc.) in knowledge engineer of Erdas Software to make your classification more accurate. If I were you I would use Ecognition for the classification.

[styfler]

Thank you very much guys

Now I'm having problems to import the results of ecognition on Arcmap.

I exported the pansharped landsat8 rgb composition from arcmap to Ecognition. I made a really nice map, but when I import the shap result to arcmap, its came with a "inconstant extent error". I've tried some projection tools, change the project at arcatalog, but the classified map doesn't show neither on the right extent or on the fly.

[Arhanghelul]

You can also export the classification from eCognition as a raster image. And then you can open the classified raster in ArcMap.

Edited June 13, 2014 by Arhanghelul

[styfler]

Well, its funny.

After almost lost all my hair, I figure that when I got the landsat 8 images on the "wrong" projection (wgs_84_zone_23N) that somehow messed my data frame extent. Since I like to work with the data frame "clipped" to the region of my interest by a polygon,and yesterday I shut of this option, today I clipped the data frame again and the shape opened "on the fly" in the right extent..... Finally.

Thanks!!

[sawyer]

I do OBIA classification of Landsat images. I tried the software ENVI and Ecognition,for extraction, but something is wrong. Water and forests are we shown the same. Can someone tell me what the problem is, which bands to use?

[meodensi]

Don't forget the TIR (Band 61, 62) from Landsat. The infrastructure is of higher temperature than the others. In reverese, water bodies are of low temperature.

 

ENVI 5.x is also able to segment the image with lots of parameters: shape, size, perimeter... have you tried?

 

Lastly, don't try to rely completely on automated solution from image processing softwares. You may think of post-processing algorithms, in which lots of eCognition's functions can be implemented by ArcGIS/ArcView/MapInfo easily.

 

Regards.