How to classify forest/land covers using Landsat imagery

[skylmla]

Dear all,

Please kindly guide me how to classify forest/land cover using Landsat imagery with ERDAS software?

Thanks in advance...

Sky

[rahmansunbeam]

Hello skylmla,

You have two kind of classification techniques, supervised and unsupervised. Supervised classification refers that the observer knows all the classes the pixels stands for. This means the pc will depend on the user defined classes for the classification. The unsupervised classification means the user do not have any idea about the class, so the pc will take control and do the classification. Another very popular technique is hybrid classification. I will tell it later.

To identify a class you'll need to collect signature points. These points are the reference points pc use for assigning certain pixel to certain class. As an example, if you are looking for bare lands in an image, you'll have to identify few bare lands by yourself (supervised classification). Then take some signature points from them. Give it to your machine. The machine will calculate the spectral values and identify similar pixels to form a new class called 'bare land'. Now there is one more thing! If this is truly a supervised classification you'll need to visit the ground and collect exact location (lat and long) to make sure you are right. This makes two signature points. Which one is accurate? use the accuracy assessment to define that.

If you are using unsupervised classification, the machine will do all these stuffs for you. You'll get signature points and classifications which the machine feels right. But, it may not be accurate in some ground.

Did you noticed that you have signature points created for the unsup class? Yes. If you use it for your supervised classification it will be hybrid classification.

All these chit chat was about classification. Now lets talk about forest and vegetation classification. You'll need NDVI indices to identify the vegetation in your image. 

[skylmla]

@rahmansunbeam

Thanks a lot for your quick response. What i am planning to do are as the following, and please help to correct and answer some of my question. I hope you don't mind...

1- download Landsat-7 (and maybe other imagery) from http://edcsns17.cr.usgs.gov/NewEarthExplorer/

  - I am not sure if the downloaded images will be ready for classification process, or I have to do Geometric and radiometric corrections and Image Enhancement? and how to do these Geometric and radiometric corrections and Image Enhancement?

2- field sampling (for creating signature points and control points)

3- Using ERDAS to classify vegetation classes.

4- Compute confusion matrix to access accuracy of the classification based on the control points.

5- Repeat step 4 and 5 in order to compare accuracy of different classification algorithm.

6- Select the most accurate classification algorithm for generating final vegetation map.

Again, please correct me and give some other tips.

Thanks

[rahmansunbeam]

Well, i dont know which are the images you'll end up from this link. Not every image here are fully clear. You may need to do few atmospheric corrections if the cloud cover is too high (i never found 0%). You may also need orthorectification. Landsat images are georeferennced, so need no geometric correction. Consider the time of the image, not every season shows similar vegetation coverage (i used to write a paper on this topic not very long ago).

Are you planning to classify the vegetation species as well?

Others are ok.

[skylmla]

If possible, we want to classify the vegetation species as well, but I am just afraid my RS skill could not reach this intend. You could give a bit explanation about classify the vegetation species?

And you are right, there is no 0% of cloud cover, so what is the maximum % we could use for classification? or Is there a way to remove cloud from the images?

One more thing, in Landsat-7 I downloaded, there are many black lines? Is this normal? and How could we remove them from the images?

Thanks,

[rahmansunbeam]

Classifying vegetation species is simple, but you can't do that with landsat. You'll need higher resolution image (ie lider, irs, and others like these).

Removing a feature from an image is impossible. All you can is to minimize its effect. Try downloading as small cloud cover as possible. Then try atmospheric correction, you'll get the closest possible result.

Did you download SLC-off images? Download SLC-on. You'll see that at the download option. Then stack them sequentially. Then see what happens.

[skylmla]

yes. I used L7 SLC-off (2003-present). Other options could be L7 SLC-on (1999-2003), but there is no recent image after 2010 or 2011?

[rahmansunbeam]

SLC-off are rubbish! You can't work with them. Recent images are not for immediate download. You'll need to send a request for them. Try this

[hide]

http://glovis.usgs.gov/

[/hide]

[skylmla]

so we could not use landsat 7 after all?

[skylmla]

the download options say "Level 1 Product ( 232.7 Mb Geotiff)/". I wonder what is Level 1 product?

[rahmansunbeam]

Yes you can. You have to consider the facts. Try TM5.

[skylmla]

There is a program called LANDSAT 7 Gap Fill in our forum at http://gisarea.com/index.php?topic=552.0.

I am not sure if this could help or not...

[rahmansunbeam]

Never tried that. If you use thematic mapper you'll find latest one.

[3dbu]

you can use lansat sensor for make this type of classification,just need control point or sample for make the signatures and make the classification ....i make many classification using this sensor , and the accuracy is accept and the data ares correct like field status.

regards

[skylmla]

@3dbu and rahmansunbeam,

Regarding the samples (both for signature and for controls), do you have formula to calculate number of samples? for example, based on each class or study area sizes or...?

[rahmansunbeam]

Number of the samples do not follow any formula. Its a trial-and-error process. But be careful when you take signatures for a class. If you take too much sample for a class and leave others for lesser sample, the whole accuracy will be biased towards that class. Try to take equal number of samples for every class. Use area fill tool. 

[3dbu]

no have any formule but just make some sample for each class used in your classification , or create a vector with this class or ID, now you can use ARCGIS Classification Toolbar and make a fast and good classification....

[marmoechlisin]

Classifying vegetation species is simple, but you can't do that with landsat. You'll need higher resolution image (ie lider, irs, and others like these).

Removing a feature from an image is impossible. All you can is to minimize its effect. Try downloading as small cloud cover as possible. Then try atmospheric correction, you'll get the closest possible result.

Did you download SLC-off images? Download SLC-on. You'll see that at the download option. Then stack them sequentially. Then see what happens.

Hi

its interesting topic, now i'm try to identify vegetation and  non-vegetation used Spot5, 5m resolution, may i know how to classifying vegetation species?

thanks

[3dbu]

for make a species classification is neccesary have sample of each species to classify...you have to get spectral invervals of each species and is necessary have sample....good luck .

[pasfans01]

just for your information if you may be not know yet,

GLCF has just released new landsat surface reflectance dataset into its portal, its been orthorectified and atmospheric corrected also, so you can use it to classify vegetation and other landcover

check it out at

http://landcover.org/data/gls_SR/

[Haruka3]

2 everybody. Anyone can answer me this problem?

I use ENVI 4.8 to classify landuse. After did it, I want to evaluate the results but I don't have enought ground truth pixels, so I use "generating random sample" to take pixels (ENVI: Classification\Post Classification\Generating random sample\Using Ground truth roi)

Random pixels appeared but they are the same location with training data. I think they must distribute over image and different place with training data.

Anyone who met same problem, can answer me?.

Thank you.

[gisadept]

2 everybody. Anyone can answer me this problem?

I use ENVI 4.8 to classify landuse. After did it, I want to evaluate the results but I don't have enought ground truth pixels, so I use "generating random sample" to take pixels (ENVI: Classification\Post Classification\Generating random sample\Using Ground truth roi)

Random pixels appeared but they are the same location with training data. I think they must distribute over image and different place with training data.

Anyone who met same problem, can answer me?.

Thank you.

I don't really understand what your reasoning is of 'not enough ground truth (GT) pixels SO generate random sample'. If you don't have enough ground truth pixels in total, you'll have to define more GT ROIs. It's the user that determines what GT is; you can't command the software: OK, here's an image, draw a sample and you use this to assess the classification accuracy. That just does not make any sense.

Or am I misinterpreting your words?

Anyway: the module you use is meant to draw a random sample from an image containing pixels that belong to ground truth ROIs. So then your GT ROIs are a pool out of which you take a sample.

[arcsump]

Hi

its interesting topic, now i'm try to identify vegetation and non-vegetation used Spot5, 5m resolution, may i know how to classifying vegetation species?

thanks

i believe for this to at least work, you must identify their spectral signatures (i.e. SS for corn, ss for coconut, etc.)

[arcsump]

2 everybody. Anyone can answer me this problem?

I use ENVI 4.8 to classify landuse. After did it, I want to evaluate the results but I don't have enought ground truth pixels, so I use "generating random sample" to take pixels (ENVI: Classification\Post Classification\Generating random sample\Using Ground truth roi)

Random pixels appeared but they are the same location with training data. I think they must distribute over image and different place with training data.

Anyone who met same problem, can answer me?.

Thank you.

You can select Regions of Interests (ROIs) yourself, assigning each class to a specific color you may wish., after which you can go with unsupervised or supervised classification.

Ihave done this before, and it works.