I would like to confirm that FragStats metrics can calculated for specific sample locations without developing values for the effective raster area. In this way only the focal window area around each sample coordinate pair would be generated. This would greatly speed up processing time and external model development through covariate selection. Once significant covariates are determined and if they are FragStats based, the user could develop full raster wide metrics for use in distributed prediction.
Yes, Fragstats has several sampling options for computing metrics within local windows. In your example, you can input your own point locations or fragstats can generate random point locations. I believe this will meet your needs.
