One of the features used in the supervised L4A processor is computed from variation of the NDVI sampled at each pixel around the current date. Specifically, the NDVI is averaged over the previous
window days, and over the following
window days. The difference between these two averages is used as a feature in the classification. So a smaller window should yield a feature that’s quicker to respond to changes, but more sensitive to noise.
I wasn’t involved when the algorithms were chosen and tuned, so I can’t say too much about the specific values of these parameters.
Generally, where can one find more information on the L4A parameter implementations (e.g. resampling_mode: “resample vs. gapfilled”)
Unfortunately, they’re not documented too well. In that specific case, it probably shouldn’t be changed. The two L4 processors use a gap-filling step to fill in the cloudy pixels by linearly interpolating between the closes (in time) valid pixels. In addition, a “temporal resampling” step is used in some cases to produce a uniformly sampled time series (to account for e.g. missing acquisitions). This parameter can enable or disable this resampling step, but it’s probably not something that should be changed.