Data & sampling
One data-rich domain, strictly paired sampling
- Products. IMERG V07 Final: IR → Uncal → Cal — three fields of one processing chain.
- Reference. GV-MRMS 30MGCP, conservatively remapped to the 0.1° IMERG grid. A reference, not truth.
- Domain. 5°×5° SE U.S., 50×50 cells, 2022–2023, half-hourly.
- Symmetric masking. MRMS quality control blanks 5.3% of pixel-times; those pixels are masked in IMERG too, so every product is summed over an identical pixel set. T = 33,769 paired times.
Without this, a product's raining area would include places the reference never had a chance to observe.
4
How it works · 1
A 3-day window: occurrence excess hidden inside an equal total
MRMS rains in 60% of frames, Cal in 78% — yet the two totals agree to 0.3%.
The budget says why: Df = +0.26 is cancelled almost exactly by
−0.05 −0.08 −0.13.
Equal totals, opposite structure.
5
How it works · 2
A single event: the fw = 1 limit
- Every frame is wet, so fw = 1 and occurrence drops out: the framework reduces to Δt·T·IwAwKw.
- Same volume (2.14×106 mm km²) from a visibly different storm: Cal is weaker (−0.27), broader (+0.18), more coupled (+0.09).
- The It–At trajectories occupy different parts of the plane.
The two windows together show which factors carry the compensation depends on the domain — and that the same algebra covers both.
6
Main result
Two years: same total, three different pathways
Shared by all three fields: rain registered more often (+0.11 to +0.17) and
27–32% weaker when it rains. Neither is removed along the chain.
They differ spatially: IR is 27% broader with MRMS-like coupling;
Uncal/Cal are area-neutral but more strongly coupled (+0.14 to +0.16).
7
Time variability
Which differences are persistent, which fluctuate
- Occurrence excess — positive in 17–22 of 24 months (p ≤ 0.002).
- Intensity deficit — negative in 23 of 24 months for every field (p < 0.001). The most robust signal in the study.
- Coupling excess — consistent for Uncal/Cal (20–21 of 24); variable for IR.
- Wet-frame area — the exception: large and positive for IR, but oscillating around zero for Uncal/Cal.
DV is the noisiest row. Its monthly median is indistinguishable from zero for every field — the total hides everything the components reveal.
10
Seasons
Area neutrality is a cancellation, not an agreement
Warm season: occurrence nearly matches; the contrast is intensity (−0.35 to −0.47) and
broader area (+0.13 to +0.43).
Cool season: IMERG rains far more often (+0.23 to +0.25), the intensity deficit halves, and Uncal/Cal become
narrower than MRMS (−0.12, −0.13).
The two-year wet-frame area neutrality of Uncal/Cal is a seasonal sign reversal — broader in warm, narrower in cool. A single aggregate number would have reported "no area difference."
11
Robustness
What survives every analysis choice — and what does not
- Robust signs across thresholds 0.1–1.0 mm hr⁻¹, windows 3 d–season, sub-domains 0.5–5°, grids 0.1–1°: occurrence excess, intensity deficit, multisatellite coupling excess.
- Scale dependent: raising the threshold shifts signal from occurrence into area — the area neutrality is specific to p₀ = 0.1.
- A process clue: coarsening to 0.5–1° removes most of the Uncal/Cal intensity deficit (−0.36 → −0.07) but not IR's (−0.26). The multisatellite deficit lives below ~0.5° — consistent with smoothing of intense cores.
Also checked: block length (1/2/5 d), coverage threshold, symmetric vs one-sided masking, co-wet frames only.
12