You can use global semivariogram analysis to determine a spatial distance. Sampling within the distance is autocorrelated. So if you want to avoid spatial autocorrelation, your ground sampling distance should be larger than it. Or you can try wavelet analysis, it will yield similar results but with lots spatial variation details.

You can also use the formula provided by "Mccoy, R. M. (Ed.) 2005. Field Methods in Remote Sensing, pp. 23 (New York: The Guildford Press)" . Using the formula, for instance, the minimal sampling plot size for Landsat TM images should be 60 m * 60 m. I would like to find the formula for you, but the book is not at hand.

GSD=P*H/F

GSD: Ground Sample Distance (cm/pixel)

P:pixel size (micron)

H: Flight Height (m)

F: Focal length (mm)

In easy way

you can use size of pixel and scale of imagery

regards

GSD = (flight Altitude x Sensor height)/ (focal length x image height &/or width). In the denominator, you calculate one time with the image height and another with the image width, then take the worst value (worst case scenario).

ground sample distance GSD = Satellite height × Instantaneous Field of View angle IFOV

More Precisely,

Let's, Sensor, (Sw x Sh) = 13.2 x 8.8 mm

Focal length, F = 8.8 mm (rear focal length, not 35 mm equivalent)

The flight height, H = 100 m

Video resolution, (imW x imH) = 4096 x 2160 (in pixels)

Then,

Ground Sampling Distance, (GSD) = (H * Sw) / (F * imW) (in cm)

Also, Image width footprint on the ground = (H * Sw) / (F ) (in meters)

GSD Calculator: https://www.3dflow.net/ground-sampling-distance-calculator/