Page 1 of 1

[SOLVED] lissage des cartes krigées

PostPosted: Wed Aug 22, 2018 10:53 pm
by Ekolleessoh
mon problème c'est le lissage de mes cartes.
j'utilise la méthode suivante prise dans le tutoriel de Didier
grid.db = kriging(data.db,grid,data.model,data_voisi_gli,radix="KM.All")
je voudrais donc obtenir des cartes bien lisse en vi=ue de leur exploitation
ci joins la carte que j'ai et qui n'est pas exploitable
[Edit: Conversion de la PJ au format PNG pour permettre de la visualiser]
esti_var1.png (102.58 KiB) Viewed 200 times

lissage des cartes krigées

PostPosted: Fri Aug 24, 2018 4:39 am
by Ekolleessoh
j'ai un problème avec mes cartes. en fait elle ne sont pas du tout lisse. j'ai éssayé d'augmenter le "radius" mais jusque la rien du tout. j'ai aussi éssayé la fonction grid.smoother sans succes ensuite krimage mais j'ai les messages d'érreurs suivantes:
The number of variables of the Data (2)
does not match the number of variables of the Model (1)
The Image Smoother is only programmed for a single variable # pour krimage et pour grid.smoother.

que faire s'il vous plait?
autre chose lors de l'exécution de la commande : kriging(data.db,grid.db,data.model,data_voisi_gli,radix="KM.All") j'ai un message du genre Error in matrix inversion (rank=8680) : Pivot #2 is null
Singular matrix
pourquoi ce message? j'ai beau retourner la question mais je ne vois pas ce qui cloche.


Re: lissage des cartes krigées

PostPosted: Fri Aug 24, 2018 8:07 am
by Fabien Ors
J'ai fusionné vos deux messages (qui traitent du même sujet) et je passe en anglais afin de permettre à la communauté d'utilisateurs de nous comprendre.

For obtaining a smoothing estimation map by kriging you should first check your moving neighborhood parameters.
Try with a unique neighborhood just for test:
Code: Select all
neigh = neigh.create(ndim = 2, type = 0)

If you have too much data samples, you should use an ellipsoid neighborhood (type = 2) with an ellipse size big enough and few sectors.

Could you plot your data location over your estimation map please ?
Code: Select all

The error regarding the matrix inversion is usually due to the presence of duplicate samples (two samples at same location) in the data set.
Please, execute the function "duplicate" before calculating your experimental variogram and its model and launching kriging.

Hope this helps.

Re: lissage des cartes krigées

PostPosted: Fri Aug 24, 2018 8:51 am
by Didier Renard
I agree with the comments of Fabien. Your problem is essentially linked to:
- the large number of data ... and there spatial density
- the use of a moving neighborhood
In addition, one can think that your data could be organized along lines (such as 2-D seismic lines).

Then I also agree that you should avoid having points too close one to the other: they are probably redundant and they will destroy the kriging system. A good recommendation is to run the 'duplicate' facility which will enable you to discard the points sitting too close to each other. You can also perform your own data selection, prior to any geostatistical procedure in order to select the GOOD samples that you ultimately want to keep.

When you have CLEANED your data set, then you can start running the estimation using geostatistical procedure such as kriging. As you have too many data, you probably have to use a moving neighborhood (the maximum number of data that can be processed per neighborhood is usually around 300: if you have less, it is recommended to use a Unique Neighborhood; otherwise, you HAVE to use a Moving Neighborhood).

The dimensioning of a Moving Neighborhood is an art which requires a lot of practice:
1) You should first enlarge the size of the neighborhood circle (or ellipse) centered on the target site and beyond which no data will be selected (the larger the better: it can even be set to the size of the whole field)
2) Then you should pay attention to the number of samples to select: this will determine the size of your kriging system. Again think to a maximum of 300 samples per neighborhood. This can be set in the parameter: Optimum number of samples (per sector).
3) If the density of samples is very irregular over the field (presence of lines of data or clusters), then it is very good habit to force the spread of the neighboring data by using a large number of "angular sectors" (you can easily take 50 sectors). Then the neighborhood consists in selecting regularly the closest samples per sector. If you select N angular sectors, make sure that the optimum number of points per sector P is selected so that N x P < 300

If the result is not yet as smooth as expected, you should try to send a picture with the data set displayed using proportional size symbols in order to let us guess:
- the density (spatial spread) of the data
- the heterogeneity of the data

Last but not least, after a correct neighborhood has been selected, the kriging procedure will provide a nice map only if the spatial model (which illustrates the continuity of the spatial characteristics of your variable) has been fitted correctly. In particular, choosing a model which is too smooth compared to the data will also tend to create strange artifacts in the resulting map.

Hope this will help