[GRASS-user] High-resolution agricultural land cover from satellite imagery

GRASS GRASS user
1

Dear GRASS users,

(This is kind of new topic to me.)

After reading this paper that addresses the mixed-pixel issue via neural networks using Landsat Thematic Mapper (TM) data:

Tatem, A. J., Lewis, H. G., Nixon, M. S. and Atkinson, P. (2003) Increasing the spatial resolution of agricultural land cover maps using a Hopfield neural network. Int Journal of Geographic Information Sciences, 17, (7), 647-672. <http://eprints.soton.ac.uk/260104/1/tatem_tgis.pdf&gt;

I have searched for GRASS documentation on image classification, particularly on land cover. The starting point is the wiki page on image classification <http://grass.osgeo.org/wiki/Image_classification&gt; as well as section 8.6 of the GRASS book ("Thematic classification of satellite data"). They both give good basic reference info, but additional pointers a welcome.

Also found some neural network work on the topic done with GRASS <http://www.ncgia.ucsb.edu/conf/SANTA_FE_CD-ROM/sf_papers/muttiah_ranjan/muttiah.html&gt;, which seems relevant but implemented in GRASS 4.1, and hence I am unsure it survived inside a GRASS module to date (at least, I could not find it).

We are targeting an agricultural area in southern Italy (several thousands hectares) for which we have full orthophoto coverage (0.5 meters resolution), and Landsat TM data can apparently be downloaded freely from <http://glcf.umd.edu/data/landsat/&gt;\. High-resolution agricultural land cover might seem overkill, but the area is highly fragmented and hence standard CORINE land cover data tend to classify most of the land as mixed types (not very helpful).

I would like to ask a general recommendation on the best way to approach an agricultural land cover task such as the one outlined above, together with possible info on previous implementation of increasing spatial resolution of agricultural land cover maps in GRASS via neural networks or other approaches.

Kind regards, thanks in advance and apologies for a long post,

Luigi

2

I wonder why you want to use the Landsat TM data for the analysis of
the high resolusion agricultural land cover.

The Landsat TM, not ETM, has very long histry and is not the High-resolution
data as for both spatial and frequential points of view.
There are so many research papers for the analysis on agricultural
land cover.

The "esa" or Italy scientists must have much information for your interest.
Are you a scientist in Italy?

We are targeting an agricultural area in southern Italy (several
thousands hectares) for which we have full orthophoto coverage (0.5
meters resolution), and Landsat TM data can apparently be downloaded
freely from <http://glcf.umd.edu/data/landsat/&gt;\. High-resolution
agricultural land cover might seem overkill, but the area is highly
fragmented and hence standard CORINE land cover data tend to classify
most of the land as mixed types (not very helpful).

I would like to ask a general recommendation on the best way to approach
an agricultural land cover task such as the one outlined above, together
with possible info on previous implementation of increasing spatial
resolution of agricultural land cover maps in GRASS via neural networks
or other approaches.

Kind regards, thanks in advance and apologies for a long post,

Luigi
_______________________________________________
grass-user mailing list
grass-user@lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/grass-user

3

Dear 山田 康晴,

Thanks for your kind reply.

On 15/05/2012 03:18, 山田 康晴 wrote:

I wonder why you want to use the Landsat TM data for the analysis of
the high resolusion agricultural land cover.

The reason is that I found that paper I cited in my previous email
(Tatem et al. 2003;
<http://eprints.soton.ac.uk/260104/1/tatem_tgis.pdf&gt;\), which described a
way to increase resolution of land cover. I thought higher resolution
would be a good thing because of the highly fragmented agricultural
landscape I was targeting (the paper by Tatem and colleagues also
analyzes an area with small-scale agriculture in Greece).

The Landsat TM, not ETM, has very long histry and is not the High-resolution
data as for both spatial and frequential points of view.
There are so many research papers for the analysis on agricultural
land cover.

I have accessed the grass-user mailing list seeking for a possible
GRASS-based approach to the task. Hence, I would be very glad if you
could point to a couple of the research papers you refer to.

The "esa" or Italy scientists must have much information for your interest.
Are you a scientist in Italy?

Yes, I am based in Italy and my background is mostly in applied ecology.
Of course people at ESA are expert in the field. My goal when accessing
this mailing list was to see if more GRASS-related info on the topic
would emerge that may benefit me and other GRASS users.

Kind regards and thank you,

Luigi

We are targeting an agricultural area in southern Italy (several
thousands hectares) for which we have full orthophoto coverage (0.5
meters resolution), and Landsat TM data can apparently be downloaded
freely from <http://glcf.umd.edu/data/landsat/&gt;\. High-resolution
agricultural land cover might seem overkill, but the area is highly
fragmented and hence standard CORINE land cover data tend to classify
most of the land as mixed types (not very helpful).

I would like to ask a general recommendation on the best way to approach
an agricultural land cover task such as the one outlined above, together
with possible info on previous implementation of increasing spatial
resolution of agricultural land cover maps in GRASS via neural networks
or other approaches.

Kind regards, thanks in advance and apologies for a long post,

Luigi
_______________________________________________
grass-user mailing list
grass-user@lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/grass-user

4

Thank you for your reply.

And your quoted research paper(by Dr.Tatem,etc.) is very informative on me.
I can find "Superresolution Mapping Using a Hopfield Neural Network
With Fused Images" by Minh Q. Nguyen, Peter M. Atkinson, and Hugh G. Lewis,
IEEE Transaction on Geoscience and Remote Sensing, vol.44, No.3, pp736-749
(2006).

But I have never done neural network analysis on remote sensing visible
band data. And I can find the function or tool of neural network analysis
only in the ENVI/IDL software package in my laboratory.

And I cannot find any tool or module on neural network analysis
in the GRASS software.

I will introduce some research paper for solving mixel problem as follows.
-------------------------------------------------------------------------------------
1. Yoshiki Yamagata and Yoshifumi Yasuoka (1996) “Unmixing wetland vegetation types by subspace method using hyperspectral CASI image”, Int. Archives of Photogrammetry and remote
Sensing, 31(７), pp.781-787

2. The VSW index method and its algorism by Yoshiki Yamagata, et al
    The Journal of the Remote Sensing Society of Japan (in Japanese), 17(4), pp.54-64(1997)

3. David A. Landgrebe (2003) "Signal Theory Methods in Multispectral Remote Sensing",
    Wiley Series in Remote Sensing book ISBN 0-471-42028-X
   There is some explanations for the hyper spectral remote sensing data analysis,
   such as the Spectral Angle Mapper(SAM). And this book is the manual of "MultiSpec",
   a kind of open source software.
   https://engineering.purdue.edu/~biehl/MultiSpec/
   When I sent my e-mail to the Purdue university developing team, Dr. David A.
  Landgrebe, professor emeritus of Purdue university, directly replied to my question.
-------------------------------------------------------------------------------------

Thank you.

(2012/05/15 17:20), Luigi Ponti wrote:

Dear 山田 康晴,

Thanks for your kind reply.

On 15/05/2012 03:18, 山田 康晴 wrote:

I wonder why you want to use the Landsat TM data for the analysis of
the high resolusion agricultural land cover.

The reason is that I found that paper I cited in my previous email
(Tatem et al. 2003;
<http://eprints.soton.ac.uk/260104/1/tatem_tgis.pdf&gt;\), which described a
way to increase resolution of land cover. I thought higher resolution
would be a good thing because of the highly fragmented agricultural
landscape I was targeting (the paper by Tatem and colleagues also
analyzes an area with small-scale agriculture in Greece).

The Landsat TM, not ETM, has very long histry and is not the High-resolution
  data as for both spatial and frequential points of view.
There are so many research papers for the analysis on agricultural
land cover.

I have accessed the grass-user mailing list seeking for a possible
GRASS-based approach to the task. Hence, I would be very glad if you
could point to a couple of the research papers you refer to.

The "esa" or Italy scientists must have much information for your interest.
Are you a scientist in Italy?

Yes, I am based in Italy and my background is mostly in applied ecology.
Of course people at ESA are expert in the field. My goal when accessing
this mailing list was to see if more GRASS-related info on the topic
would emerge that may benefit me and other GRASS users.

Kind regards and thank you,

Luigi

We are targeting an agricultural area in southern Italy (several
thousands hectares) for which we have full orthophoto coverage (0.5
meters resolution), and Landsat TM data can apparently be downloaded
freely from<http://glcf.umd.edu/data/landsat/&gt;\. High-resolution
agricultural land cover might seem overkill, but the area is highly
fragmented and hence standard CORINE land cover data tend to classify
most of the land as mixed types (not very helpful).

I would like to ask a general recommendation on the best way to approach
an agricultural land cover task such as the one outlined above, together
with possible info on previous implementation of increasing spatial
resolution of agricultural land cover maps in GRASS via neural networks
or other approaches.

Kind regards, thanks in advance and apologies for a long post,

Luigi
_______________________________________________
grass-user mailing list
grass-user@lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/grass-user

--
-------------------------------------------
Yasuharu Yamada
Chief Researcher,
Research Project for Resources Information Technology,
NIRE, NARO Japan
yamaday@affrc.go.jp
http://nkk.naro.affrc.go.jp/
----------------------------------------------------------------------------------
This e-mail is intended only for the designated recipient(s) and may
contain confidential information. Please don't forward the message
to the other person(s) or place(s) without prior notice.

5

On Wed, May 16, 2012 at 11:01 AM, YAMADA,Yasuharu <yamaday@affrc.go.jp> wrote:
...

And I cannot find any tool or module on neural network analysis
in the GRASS software.

You may use neural networks in R (using the GRASS-R interface).
For a general overview of the integration, see
http://grass.osgeo.org/wiki/R_statistics

I will introduce some research paper for solving mixel problem as follows.

...

3. David A. Landgrebe (2003) "Signal Theory Methods in Multispectral Remote Sensing",
Wiley Series in Remote Sensing book ISBN 0-471-42028-X
There is some explanations for the hyper spectral remote sensing data analysis,
such as the Spectral Angle Mapper(SAM). And this book is the manual of "MultiSpec",
a kind of open source software.
https://engineering.purdue.edu/~biehl/MultiSpec/

As my master thesis, I wrote these two modules:

i.spec.sam: Spectral Angle mapping
http://grass.osgeo.org/wiki/Addons#i.spec.sam

i.spec.unmix: Spectral unmixing
http://grass.osgeo.org/wiki/Addons#i.spec.unmix

The latter needs proper update to GRASS 6 but that should not be too hard.

Markus

6

Dear Markus

On 16/05/2012 14:19, Markus Neteler wrote:

...
You may use neural networks in R (using the GRASS-R interface).
For a general overview of the integration, see
http://grass.osgeo.org/wiki/R_statistics

Right. Sometimes I forget the secret weapon...

...
As my master thesis, I wrote these two modules:

i.spec.sam: Spectral Angle mapping
http://grass.osgeo.org/wiki/Addons#i.spec.sam

i.spec.unmix: Spectral unmixing
http://grass.osgeo.org/wiki/Addons#i.spec.unmix

The latter needs proper update to GRASS 6 but that should not be too hard.

I had a feeling I was missing something important.

In the description of the two add-on modules, I found reference to two very interesting papers that triggered some more search. In particular, this conference paper by Stabile et al. (2009) on Fusion of High-resolution Aerial Orthophoto with LandSat TM Image for Improved Object-based Land-use Classification, uses the same data layers I may access (orthophoto + Landsat TM):
http://www.a-a-r-s.org/acrs/proceeding/ACRS2009/Papers/Oral%20Presentation/TS12-05.pdf

based on which, I wonder if it would make any sense to perform i.fusion e.g.

i.fusion.brovey -l ms1=lsat7_2002_20 ms2=lsat7_2002_40 \
                    ms3=lsat7_2002_50 pan=ortho_photo_rgb_composite \
                    outputprefix=brovey

and then use i.smap after i.gensigset (the use of group= and subgroup= parameters in both modules is a bit unclear to me) to classify the map using areas with known land cover class identified via the orthophoto itself or field survey.

Kind regards and thank you (and pardon my ignorance),

Luigi

7

In the description of the two add-on modules, I found reference to two very interesting papers that triggered some more search. In particular, this conference paper by Stabile et
al. (2009) on Fusion of High-resolution Aerial Orthophoto with LandSat TM Image for Improved Object-based Land-use Classification, uses the same data layers I may access
(orthophoto + Landsat TM):
http://www.a-a-r-s.org/acrs/proceeding/ACRS2009/Papers/Oral%20Presentation/TS12-05.pdf

I can find the function of Self-Organizing Map Classifier using
Neural Networks in the Hyperspectral Analysis module of TNTmips software
of Microimages,Inc. in my laboratory.
But it is for the hyper-spectral (maybe over 300 bands) data such as CASI,
not for LANDSAT 7-bands data.

And I wonder whether the fusion data of aerial photo and satellite low resolution
data can be used for computational classification.
The fusion data (panchromatic sharpened data) is suitable for the photointerpretation.

But thank you for your introduction of some research papers.
This is very informative for me.

--
-------------------------------------------
Yasuharu Yamada
Chief Researcher,
Research Project for Resources Information Technology,
NIRE, NARO Japan
yamaday@affrc.go.jp
http://nkk.naro.affrc.go.jp/
----------------------------------------------------------------------------------