spatial statistics for spectral indices and land cover composition

[olivergunawan]

Hi,

I have a created a series of spectral indices for a sample of pixels. I have also determine land cover proportions from arial photography (grass, trees, water, built and bare earth)

I am looking for statistical methods for analysing the indices against land cover proportions. Currently, I have completed descriptive statistics and as the datasets are non-parametric, I have used Mann-Whitney and Kruskal-Wallis tests to compare medians.

Does anyone have any suggetions for follow-up methods for deeper analysis?

Thanks
Oliver

[JeffreyEvans]

I am unclear on what you mean by the data being non-parametric. In statistics non-parametric refers to a statistic and not a data distribution. If you have your data organized by classes you can apply the M statistic to assess spectral separability. The M statistic is a ratio between the differences in the mean and standard deviations between two classes for each band and expects normality in the data distribution. The expectation for the M-statistic is: M<1 poor separability, M>1 good separability.  

M[1...n]=( mean(class1) �?? mean(class2) ) / ( sd(class1) �?? sd(class2) ) 

If your data is highly skewed then you can perform a transformation (i.e., natural log) to make the data more normally distributed. If you do not have your "land cover proportions" classified you could assign classes based on percentile breaks. If you would like to keep your data "continuous" a simple log transformed OLS regression or ANOVA would likely give you what you are looking for. In a regression low R^2 would indicate separability with a higher correlation indicating non-independence in cover proportions. I would however, expect that, even with a transformation applied, the relationship will be non-linear. Because of this it is important to look at scatter and residuals plots to revel non-linear relationships.