Per-station validation

Every product scored against the independent BMKG gauges, whole-record (the data behind the thesis Taylor diagram). Switch between individual stations - one dot per station for a chosen stage - and a summary that greys the stations and drops in one pooled dot per product: the raw satellites (IMERG-L, IMERG-F), the calibration target (CPC-UNI), and the three correction stages.

Taylor diagram

Angle encodes correlation (the rays and outer labels), radius the standard-deviation ratio (the grey arcs, labelled along the diagonal). The dashed arc is the reference SD = 1; the black star (SD 1, r 1) is a perfect match - closer is better.

Correlation and standard-deviation ratio are whole-record statistics against BMKG (the Taylor source); the detection scores (POD, CSI, FAR, KS) in the individual card are the dekad-pooled medians.

Seasonal stability

Dekadal pooling hides the within-year cycle. Re-aggregated to calendar months - the per-pixel spatial median in-sample against CPC-UNI - the four representative metrics show the correction holds across Indonesia's wet and dry seasons.

Finding Every metric is flat across the seasonal cycle: LSEQM+DL's SDR holds within to all year and the stage ordering never flips. RMSE tracks rainfall magnitude - peak in the DJF wet season, trough in August - not method skill. The pooled headline is not driven by a handful of dekads.

Monthly values are the per-pixel spatial median in-sample against CPC-UNI, averaging the three dekads of each month. The dashed line is the target (SDR = 1) or the 5% KS threshold. Note the KS panel: the corrected stages sit at the test floor against CPC in-sample, the flip side of their strong KS pass against the independent BMKG gauges (out-of-sample vs in-sample). LS reads higher here only because it changes the distribution least.