Clouds, Gaps, and Shadows: Where Orchard Time Series Actually Break

The most impressive model cannot recover information that was never observed. In orchard monitoring, a clean-looking time series may contain thin cloud, cloud shadow, edge artefacts, saturated pixels, or a long gap that has been hidden by a smooth interpolation. Treating every date as equally trustworthy creates a false sense of continuity, especially in regions where the very season that matters most is also the season with the worst cloud cover.

A practical Sentinel-2 workflow therefore begins with surface reflectance, scene classification, cloud probability, and a record of how much of the field survived the mask. The Earth Engine cloud-probability collection provides a per-pixel probability and documents how it can be joined to surface-reflectance imagery; Copernicus Level-2A products also include the SCL layer and atmospheric-correction outputs. After masking, a project can use ten-day composites, temporal interpolation, or Savitzky–Golay smoothing, but each choice should preserve the distinction between an observed value and an estimated value. The 2026 almond-bloom study is a useful example because it aggregated indices at orchard level, smoothed the resulting curves, and still validated the timing against time-lapse cameras. See Copernicus Level-2A documentation, S2 Cloud Probability, and Lauterman et al..

When the record is too weak, silence is better than a confident-looking map. A production result should say how many valid observations were available, which part of the field was usable, how much of the curve was interpolated, and whether the comparison window is long enough to support a change claim. Sentinel-1 can help during optical gaps, but fusing radar and optical data introduces its own calibration and interpretation problems. The engineering goal is not to eliminate uncertainty; it is to make uncertainty visible before it turns into a field visit that cannot be defended.