Sensitivity to the three configuration parameters

Three parameters were set by design judgement, not formal optimisation: the DL blend weight (α = 0.70), the GPD tail threshold (80th percentile), and the density saturation count (2). A Bali sweep varies each around its default while holding the others fixed. The result is both reassuring and a little humbling - the daily correlation barely moves.

The shaded band is the full sweep envelope; every curve stays inside it, and each red star marks the operating default. Hover any point for its full metric set. There is no sharp optimum to miss - the response surface is flat.

The full sweep

All fifteen settings, pooled per-pixel over the Bali subdomain. The operating default of each parameter is shaded. Pearson r holds in [, ], but the distribution metrics do respond - the GPD threshold drives the most movement (SDR 0.923 → 0.867 across its range), which is why it is the one setting a downstream user would most need to revisit.

Why this is the expected result, not a lucky one

The parameters shape how the marginal correction is applied, but none of them add day-to-day timing information. Pearson r is bounded by the satellite retrieval (the r ≈ 0.34 ceiling), so tuning α, the GPD threshold, or the saturation count moves skill only within the noise. Formal optimisation would not lift the ceiling - that needs the calendar-window fix or a different class of method.

The one place a parameter matters more than this chart suggests is spatial: the density saturation count barely moves pooled r, but it decides where the CNN is allowed to act. Invariance in the pooled number is not the same as invariance in the map. The thesis flags all three as reasonable-but-unoptimised settings, and this sweep is the evidence that the choice is safe within its range.