Re: AI-GEOSTATS: Simulation and trends

[Adrian Martínez Vargas]

Chris   As have told you Isobel you can loos de main information that gave de SGS, that is the local variability. You most be careful of how mush spaced is your data, the SGS can be in occasions les accurate than the turning band (with large quantity of random bands) and other kriging methods as consequence of increasing of error associated to the sequential approach. For test the simulations you can simulate with few variants of the inner kriging methods of the SGS in know points extracted randomly from the data, if your software don’t have implemented the option of Jackknife you can migrate de points to the closer nodes of a dense grid, and compare it mean with the true value. In any way the most accurate methods must be SGS with kriging with local mean and ordinary kriging (remember the advise of Isobel about the universal Kriging)   I hop it help you   King Regards Adrian Martínez Vargas ISMM, las Coloradas s/n Moa, Holguín, Cuba CP 83329 http://www.geocities.com/adriangeologo/adrian.html ----- Original Message ----- From: Chris Lloyd To: ai-geostats Sent: Wednesday, August 06, 2003 12:27 PM Subject: AI-GEOSTATS: Simulation and trends Hello,   I am currently using sequential Gaussian simulation (SGS) to generate microtopographic soil surfaces from sparse data. There are nearly 16000 observations,  so I’m using a small search neighbourhood (e.g., 16 observations). The mean of the variable I’m concerned with (heights) increases systematically from the top to the bottom of the data set and R squared for a fitted first order polynomial is 0.885. An obvious choice is to detrend the data, use SGS based on simple kriging and then add the trend back. An alternative might be to use the fitted trend to define the locally-varying mean and apply SGS based on simple kriging with locally-varying means (rather than taking the mean of the residuals as the constant mean and applying standard simple kriging). I suspect that ordinary kriging (using a power model fitted to the raw variogram) would result in predictions as accurate as those obtained through detrending in some way, but given the trend is so obvious I don’t want to ignore it. I am aware that there is a lot of relevant work in the literature about the application of these approaches in the context of kriging. However, I would be interested in details of any case studies that have dealt with large scale trends in a simulation context. I would also be interested in the views of list members about the approaches I’ve mentioned or any others that may be appropriate.   Many thanks in advance,   Chris Lloyd