Hello,
I work as a student of a master’s degree in science, and I work on the detection of mowing events by satellite data with the use of pixels.
I read a blog by Matic Lubej about using the pixel as a mowing marker. However, I had a few questions, notably about the number of pixels necessary to define a mowing event:
You define an optimal number of pixels to detect a mowing event with 8 pixels for large plots and 3 for small ones. How statistically is 8 pixels better, do you have bibliographic sources? have you published an article on this?
Welcome to the SH Forum, thanks for asking your question, I’ll try to answer it.
You define an optimal number of pixels to detect a mowing event with 8 pixels for large plots and 3 for small ones. How statistically is 8 pixels better, do you have bibliographic sources? have you published an article on this?
The decisions we made were based on qualitative results on real data, so unfortunately there was no statistical study performed which would make the answer more clear and useful for academic purposes. The main idea was that even partial mowing is done in patches, so we had to define an area which would be significant enough to be considered as partial mowing and not a noisy fluctuation. The additional constraint we added here was that the pixel-level detections needed to be spatially connected.
How do you define a small and a large plot?
This is again a bit qualitative, but the decision is a result of multiple factors, such as the Sentinel-2 resolution (10 m), average field sizes in the area-of-interest, and the amount of noise in the observed features (in this case NDVI). Parcels which are long but narrow are also considered small, if their width is on the scale of a few Sentinel-2 pixels or fewer (regardless of their length).
So, to conclude, there is unfortunately not a lot of concrete information which could be used for strict academic purposes, we tend to focus more on knowledge sharing and story telling, and our solutions being good enough for the task at hand.