topic Re: geostatistical analyst-kriging in ArcGIS GeoStatistical Analyst Questions

geostatistical analyst-kriging

juliaDa — Wed, 04 Feb 2015 19:04:23 GMT

Hi all,

I am using IDW and Kriging in ArcGIS to explore the distribution of forest carbon stock using 183 ground sample points. The data is not normal and there is trend, so I have transformed the data. However after running kriging, the cross-validation result shows root mean square standardized: 0.98, root mean square and average standardized error: 127.7 and 128.89 respectively and the main problem is the predicted and error plot is not 1:1. Can anyone please suggest what can be the issue and how can I address this. Any helpful suggestions will be highly appreciated.

Thanks in advance!

Re: geostatistical analyst-kriging

DanPatterson_Retired — Wed, 04 Feb 2015 19:20:42 GMT

Ok they are the numbers...what is the area like from a qualitative perspective?

Re: geostatistical analyst-kriging

juliaDa — Wed, 04 Feb 2015 19:42:17 GMT

Dan thank you for the reply

Yes the sample points are numbers, the study area is watershed of about 8000 ha, and with wide elevation range from low mid 200 to high 1900 m. The main landuse in this area are forest and agriculture, there are some agricultural patches in between the forest area where there is no sample points.

Re: geostatistical analyst-kriging

DanPatterson_Retired — Wed, 04 Feb 2015 19:52:25 GMT

Before getting into the spatial interpolation methods, did inferential tests between landscapes yield any valuable information? Had you considered partitioning the landscape into discrete areas prior to kriging? since I am not sure why you would do that if there are inherent differences in the landscape to begin with (ie forest vs agri. low vs high elevations). I am just trying to get a sense why kriging is being applied to the whole area?

Re: geostatistical analyst-kriging

juliaDa — Wed, 04 Feb 2015 20:45:43 GMT

Yes i have thought of partitioning but in certain region there are not enough sample points. I am trying to look at the spatial distribution of forest carbon in whole watershed, so I thought of using kriging in whole area, and also based on some literature review. I couldn't think of any other method, any suggestion!

Re: geostatistical analyst-kriging

DanPatterson_Retired — Wed, 04 Feb 2015 23:17:01 GMT

Julia...a lack of data points in one area isn't a reason to include it in the whole area...that fact alone may be masking the true result... in otherwords, including that area may make matters worse.. If you have true barriers in the terrain, consider spline with barriers for a quick look. Partition into different areas and do your kriging...do some inferential tests of difference/association on your raw data partitioned off into different areas divided by cover type and elevation class. I guess my big recommendation is don't throw all your efforts into one test/procedure then wonder why things aren't working out as perceived. There may be many reasons for results for a large area may not yield the same results when the area is partitioned. Good luck

Re: geostatistical analyst-kriging

juliaDa — Thu, 05 Feb 2015 00:45:47 GMT

Thank you very much Dan! I will explore other options including partitioning, hope that will get better result. Again thanks!

Re: geostatistical analyst-kriging

SteveLynch — Thu, 05 Feb 2015 00:47:59 GMT

Julia

Are you using 10.1 or later, if so then look at Emperical Bayesian Kriging

Also, could you include a screenshot of the semivariogram that you are using.

-Steve

Re: geostatistical analyst-kriging

juliaDa — Thu, 05 Feb 2015 01:52:37 GMT

Steve,

I am using 10 but recently also trying EBK in 10.1. I have attached the screenshot of the semivariogram from EBK after transformation.

Re: geostatistical analyst-kriging

SteveLynch — Thu, 05 Feb 2015 02:42:50 GMT

Julia

That is great, what parameters are you using?

I.e what transformation, which model, what subset size, etc.

-Steve

Re: geostatistical analyst-kriging

juliaDa — Thu, 05 Feb 2015 03:34:01 GMT

Steve, below is the detail. I played around with different model but this is the only model, where root mean square and average standardized error are equal however the predicted plot is not 1:1.

Re: geostatistical analyst-kriging

SteveLynch — Thu, 05 Feb 2015 04:48:14 GMT

Try a smaller subset size and increase the overlap factor to 2 or 3

Re: geostatistical analyst-kriging

juliaDa — Thu, 05 Feb 2015 05:18:53 GMT

Hi Steve,

Thank you! I have tried with different subset size and factor but the result is not good. I have attached the measured and predicted value plot.

Re: geostatistical analyst-kriging

SteveLynch — Thu, 05 Feb 2015 15:37:24 GMT

Julia

Do you have any other information that might help explain the variable that you are trying to interpolate, for example, elevation, soil depth or any other continuous variable?

If so you might want to look at cokriging.

-Steve

Re: geostatistical analyst-kriging

DanPatterson_Retired — Thu, 05 Feb 2015 15:50:32 GMT

As I suggested previously...just leaving out the spatial stuff...examine any relationships/differences before looking for spatial patterns or trying to interpolate values between field observations.

Re: geostatistical analyst-kriging

juliaDa — Thu, 05 Feb 2015 20:36:09 GMT

HI all,

Thank you for all your great suggestion. I am working on elevation, soil texture and forest type as covariates, but couldn't figure out how to incorporate nominal data. I worked with other ratio covariates: population, ndvi but did not get any good result.

As Dan suggested I did some analysis in SPSS but there is no significant relation with any of these independent variables and the correlation is ow <5. I am not good in statistics, I only know basics but the descriptive result shows that standard deviation of the forest carbon stock is very high compared to mean, does that effect the result?

Re: geostatistical analyst-kriging

DanPatterson_Retired — Thu, 05 Feb 2015 21:06:38 GMT

soil texture and forest type are nominal data...ie data which are classed so for example to simplify greatly without seeing your data you could perform a Chi-Squared test of classified carbon stock versus soil type.

For the benefit of others, in a greatly simplified form... data can be categorized as either: nominal (classed data), ordinal (some inherent rank between observations), interval (equal spacing between values, arbitrary zero), ratio ( interval with true zero point). There are many tests of association/difference/correlation that one can use to test for significance with data...I always recommend that these be explored before one adds the spatial dimension in...but that is my opinion. For statistical discussions I usually turn to content on Cross Validated as Stack.Exchange site devoted to data and data analysis...I include interpolation in the category of statistical analysis since at least one level, you are looking for the degree of association between observation points. Others view interpolation in your context as non-statistical but I will not quibble with their opinions here. So the advice I was trying to give you is ... understand your data, before you add the dimension of trying to fill in spatial gaps in your observations.