4 replies to this topic

[dav]

Hey,

I'm trying to derive tree cover from a topographical map (unfortunately I've got no appropriate Landsat or other imagery for the area (it's hard to separate tree cover and fields on summer imagery) for deriving tree cover from satellite imagery). Unforunately cluster processing with diverse programs (as IDRISI or SAGA) haven't given good results. Therefore I installed the trial version of Avenzas Photohsop Addon and chose the tree cover simply with the magic wand tool. This is working quite well, but unfortunately contour lines, text and grid line are leaving white spaces.
In a GIS I could probably use the buffer function, no? Is there any possibility in Photoshop to solve this? I've tried Gaussian blur and other filters for filling the spaces (after that once again use the magic wand tool, inverting the selection and refilling the selection with only one color). Results are not bad, but in some cases the tree areas are growing too much into all directions and do not only close the white spaces within the area.

Do you have any ideas how to solve that problem?

Attached Files

•  example_trees.jpg   628.1K   21 downloads

[NicKV]

I'm trying to derive tree cover from a topographical map (unfortunately I've got no appropriate Landsat or other imagery for the area (it's hard to separate tree cover and fields on summer imagery) for deriving tree cover from satellite imagery). he area. ....

...Do you have any ideas how to solve that problem?

I rather doubt that you cannot distinguish forested areas from grassland. I suspect you are/were not using the right Landsat spectral band combination and/or the right software for the job.

[Clark Geomatics]

Try looking for a Landsat 7 or L4/5 scene (depending on the resolution you need) in the early spring or late fall (ie when there is a good distinction between trees and fields) and use the following band combination for your RGB channels (in Photoshop or any decent image processing tool) - R=band 7, G=band 4, B=band 2. Because band 7 reads in the near-infrared spectrum, it is great for distinguishing productive/non-productive vegetation. Lots of literature out there on the topic.

Good luck.

[dav]

hi,

thanks for your advices, but I already tried a lot, and deriving forest areas is only a minor part of my work (rather a visual thing than a scientific one). My research area is lacking a homogenous Landsat coverage in the important seasons (most were recorded in summer) and I unfortunately need a full coverage and not only a partly one (research area stretches over 250x250 km²). Anyway thanks for the information with the band combination, I will try it (but the NDVI raster also provided good results). Even winter imagery would fit my needs, because the research area is fully covered by a small snow pack in the winter months. Trees appear black on normal RGB and are in strong contrast to the surrouding white colors.
So...the method with the topographical maps appeared most fruitful to me. Clustering Sat. images will, according to my experiences until now, also need a lot of post editing (removing non-forest structures vom the selection, etc.). I have done this for a minor part of my research area with ASTER imagery and it worked well. But it would be too much work doing this for the whole area.

By the way, another question: Is there any rule for defining the number of cluster classes, which should be generated, for special tasks? This is always very hard to decide for me. I always try different settings, from 5 to 12 classes in order to just get the forest areas seperated. Is it a try and error thing, or are there useful rules for defining the number for such easy tasks as mine?

Thanks very much

dav

[NicKV]

To do this rigorously so you can have confidence in the results...

You will need to obtain othorectified remote sensing data that has been subjected to minimal radiometric processing, together with relevant metadata about the collection of the data by the sensor. For Landsat and ASTER and ALI data, doing this this requires that you need to obtain data supplied as HDF EOS files, not as geotiffs which carry none of the necessary metadata.

Then the pixel DN values of previously orthorectified data should first be converted to radiance-at-the-sensor values and these then transformed to target (pseudo-)radiance values by subjecting the radiance at sensor values to an atmospheric correction procedure. The resulting radiance at the surface values then need to be converted to target (pseudo-)reflectance values by performing topographic correction using a DEM, to correct for topography-caused illumination variation. The objective of the latter is to minimise the effects of shadowing, which would otherwise confuse analytical algorithms.

Photoshop cannot be used for this and in general, is not a suitable tool for remote sensing data analysis as it is not created for this purpose. It does not have the processing, analytical and classification algorithms to distinguish target types. Classification is not the same as pixel selection using the pixel selection tools that may be provided in image editing software. RS data analysis should be performed on data and not on a visualisation of the data as an image. This means performing the data analysis in pixel reflectance value information space, not geographic of an image.

The more spectral bands you use in the analysis the greater the likelihood you will be able to distinguish target types and the greater the confidence you can have that you have distinguished between them correctly.