Observation Diagnostics Overview
The observation diagnostics subsystem processes IODA files and generates a comprehensive suite of observation‑space diagnostics, including statistics, histograms, QC summaries, and ATMS‑specific channel and scan‑position diagnostics. These tools quantify how well the background and analysis fields fit the observations and complement the increment and spectral diagnostics. For the mathematical formulation and example figures, see Diagnostics Overview.
Subsystem Components
The observation diagnostics subsystem consists of the following components:
ObsDiagnostic — core engine for computing O–B and O–A departures, bias, RMS, normalized RMS, and QC‑filtered statistics
IODA loaders — readers for satellite and conventional observation files
Scalar histograms — e.g., temperature, humidity, pressure
Vector histograms — e.g., wind components (u, v)
ATMS diagnostics — channel‑wise extended RMS statistics and scan‑position bias checks
QC summaries — counts and statistics for QC‑passed and QC‑failed observations
Plotting utilities — generation of extended RMS, histogram, and scan‑position figures
Innovation‑Space Error Diagnostics
In addition to standard O–B/O–A statistics, the diagnostics subsystem
includes an innovation‑space error diagnostic based on the Desroziers
method. This tool computes innovation variance, estimated
observation‑error variance, and recommended R‑scaling factors using only
OMB and OMA. It is implemented in
ufs_da_diagnostics/obs/innovation_br_check.py and is intended for
routine monitoring of observation‑error consistency and tuning.
The diagnostic reports:
Sd = E[OMB²] — innovation variance
R_est = E[OMA × OMB] — Desroziers estimate of true observation‑error variance
Sd/R — innovation chi‑square proxy
R_est/R — variance‑scaling factor
HBHᵀ = Sd − R_est — background‑error contribution
scale_R = R_est / R — recommended R multiplier
infl_chi — standard‑deviation inflation needed to reach a target chi‑square
These quantities provide a lightweight, observation‑space method for evaluating the consistency of assumed observation‑error variances.
Workflow
A typical observation‑diagnostic workflow consists of:
Load IODA file Read observation values, metadata, and QC flags.
Load H(x) files Load background and analysis model equivalents for computing O–B and O–A departures.
Compute statistics Compute bias, RMS, normalized RMS, bias‑corrected RMS, and QC‑filtered statistics for each variable and channel.
Generate plots Produce histograms, extended RMS figures, scan‑position diagnostics, and QC summaries.
Save outputs Write statistics tables and figures to the output directory.