Hi Shariful

Have you considered developing a variogram and using a distance slightly beyond the range? 

All the best

Todd. 

The spatial auto-correlation diagnostic, Moran's I is developed for use in areal data (i.e., polygons). This is to partially to respond to limits of polygon data where the distribution of values within a polygon are unknown. Since you have point data, you can generate variables to "adjust for" spatial autocorrelation, if it exists. You can include the latitude and longitude as a latent variable. You can generate a distance to nearest neighbor variable, or distance to nearest neighbor with a characteristic of interest. To check and calculate Moran's I of your point data, I usually convert it to polygons using a Thiessen polygon.   

I am not sure what is the purpose of your analysis, but a more modern approach will be to factor the auto-correlation in your model (auto-logistic models) and estimate the parameters. It most cases it is a bad idea to throw away data. Your first instinct could be to calibrate your logistic model without considering auto-correlation and examine the residuals. If your residuals do show a spatial auto-correlation, you must extent your model to include auto-correlation by moving to auto-logistic models.

R has packages that implement it.

http://www.whoi.edu/cms/files/DormannEcography30_57164.pdf

https://booksite.elsevier.com/brochures/hugy/SampleContent/Spatial-Autocorrelation.pdf

https://www.utdallas.edu/~dag054000/Taiwan_lectures/background%20readings/ESM-2005.pdf

http://www.nat-hazards-earth-syst-sci.net/14/259/2014/nhess-14-259-2014.pdf

Perform the Auto-Correlation Function  (ACF) through SPSS software. Remove trend,seasonality and autoregression (ARIMA). Then perform regression. 

I have the same issue. Selected random samples from aphrohyte pixels for downscaling. But the sample points are autocorrelated. Please help