The CNN refinement is not applied uniformly. A confidence field, built from the smoothed density of the BMKG station network and used as a proxy for where the network contributing to CPC-UNI is dense, gates the blend: where gauges are dense the CNN contributes, where they are sparse the product falls back to pure LSEQM. This keeps the data-hungry step honest about where it has support.
Confidence over the 0.1° land grid (Gaussian-smoothed, σ ≈ km); the BMKG station locations that define it are overlaid as white dots. The network is heavily concentrated in Java - so is the confidence.
Max confidence
of 1.0 (dense Java)
Max DL weight
% where densest
Mean DL weight
% over all land
Land the CNN touches
% at a weight above 0.03%
The mask never reaches full confidence: even the densest cell tops out at , so the CNN's share peaks near % in Java and averages just % across Indonesia. The refinement is real but deliberately restrained.
The blend weight is a straight line in confidence:
effective α = 1 − confidence × (1 − ) → CNN weight = confidence ×
Confidence 0 gives α = 1.0 (pure LSEQM, no CNN); confidence 1 would give α = (30% CNN). The dashed line marks the mask's actual ceiling. The base α of is itself a judgement setting rather than an optimised one - like the saturation count below, it is one of the sensitivity parameters.
Confidence saturates at DENSITY_SATURATION_COUNT smoothed stations per cell (default ). That threshold was set by judgement, not optimised - it is one of the sensitivity parameters. Lowering it lets sparser networks reach full confidence, activating more CNN; raising it is more conservative. Drag it to see the effect on the same mask.
Mean CNN weight
% over all land
Max CNN weight
% in the densest cell
Land with CNN ≥ 10%
%
At the default of , the mean weight is % and almost no land clears a 10% CNN share - the conservative choice the thesis flags as a limitation. Dropping the count toward 1 concentrates more correction into the same Java-centred pattern rather than spreading it east, because the underlying gauge network, not the threshold, is what is sparse.