# Kriging to calculate forest biomass?

I am trying to create biomass map using Kriging from ground sample point.

I have about 150 and 53 point plot for dense and sparse strata.

The data is not normal so I have transformed it.

After running the kriging, the predicted and the measured line does not look right, also the QQ plot between normal value and standardized error is not normally distributed?

What are the options to address this issue and how can I figure out which kind of kriging is best to calculate biomass? • Perhaps it will help others if you mention what software you are using. Also explaining what you mean by "does not look right". ;) – lynxlynxlynx Jan 9 '15 at 23:41
• I am using ArcGIS and I am very new to this tool. I am trying to teach myself. The predicted and measured is not 1:1, I have attached the output. Any suggestion will be appreciated. – Julia Jan 15 '15 at 17:14
• I would highly encourage you to explore a different methodology. Forest biomass is, by nature, a highly nonstationary process thus, violating Kriging assumptions. Besides, biomass is not a purely spatial process and requires covariates to model correctly. There is a very good reason that you do not see this methodology applied in the forest mensuration literature. I would note that the Meng et al., (2009) paper uses Landsat spectral data as covariates to model a random field using various geostatistical approaches and this just not an adopted method in forest inventory efforts. – Jeffrey Evans Apr 1 '15 at 15:55

One possibility underlying the poor kriging performance could be the field plots (location and sampling intensity) failing to capture the spatial autocorrelation (or spatial dependence) in the data.

It would be necessary to observe how well the theoretical semivariogram fitted to the data (experimental semivariogram) (Figure 1). If it is the case, one possibility is to try other type of theoretical semivariograms (it seems you used the type "spherical", but there are others: Gaussian, exponential, circular, etc). Figure 1. Illustration of semivariogram parameters: sill, nugget and range (A). Example of experimental variogram and (theoretical) spherical semivariogram (B). Source: adapted from Sanz et al. (2012).

The best kriging method depends on the nature of the variable which is being studied and the type of auxiliary data available. For example: if the data is stationary (i.e., it has a constant mean), simple kriging (known mean) and ordinary kriging (unknown mean) are suitable options. On the other hand, if data is non-stationary, one option can be universal kriging. Those are types of univariate kriging.

An alternative approach would be the multivariate kriging (for example: co-kriging or regression kriging). Such methods use information from auxiliary data to enhance the capacity of spatial modelling. In the case of forest biomass, examples of auxiliary data (and auxiliary variable) are: satellite imagery (NDVI) and LiDAR (height percentiles).

The regression kriging technique for example, have one advantage which is to perform better predictions outside the sample (extrapolation), because part of the model will depend only on the relationship between the response variable and the auxiliary variable (i.e., it will not be entirely dependent on spatial variation of the data).