Paired Sensor-Ablation Supplement

Full description

This package supports the paired full multi-sensor versus Himawari-only experiment reported in “Persistent Wildfire Detection under Observation Gaps Using Multi-Sensor Thermal Data Assimilation.”. This data is to be used to gether with the MATLAB script in https://github.com/chulund/datafusion

Contents

• run_sensor_ablation.m: analysis-only MATLAB runner with smoke, full, and summary modes.
• 11_sensor_ablation_per_case.csv: validated metrics for all 1,406 pixel-cases.
• 11_sensor_ablation_summary.csv: per-event and pooled detection summaries.
• 11_sensor_ablation_metadata.json: experimental controls, metric definitions, and covariance-reconstruction note.
• tbl08_sensor_ablation.tex: generated pooled manuscript table.

Experimental definitions

The full configuration assimilates Himawari-9, GeoKompsat-2A, and Sentinel-3 observations. The Himawari-only configuration removes GeoKompsat-2A and Sentinel-3 observations in memory while retaining the same cached 10-minute timeline, Himawari observations, DTC priors, quality controls, Kalman parameters, and detector.

A Himawari-degraded slot is one where the Himawari MIR observation is missing, cloud-masked, or colder than the DTC prior by at least the configured cold-guard threshold. This is distinct from the manuscript's sparse-support gap definition, which identifies slots in which at most one of the three MIR sensors has a valid observation.

The 100,070 result is the number of Himawari-degraded slots receiving at least one accepted live measurement update in the full configuration. The GeoKompsat-2A and Sentinel-3 counts are the timeline slots in which each sensor contributed an accepted update across the complete 7,848,826-slot timeline; these counts are not additive to the degraded-slot total.

Live-update, forecast-only, and posterior-covariance metrics use the MIR diagnostics. Finite-background availability is checked jointly for MIR and TIR. Posterior covariance omitted from production diagnostics during forecast-only intervals is reconstructed using the exact scalar Riccati recurrence and must match every recorded finite diagnostic within 1e-10.

Validation and reproduction

The runner aborts unless:

• the full rerun reproduces cached MIR and TIR backgrounds and detection flags;
• the Himawari-only arm has zero GeoKompsat-2A and Sentinel-3 update slots and Kalman gain;
• reconstructed posterior covariance agrees with all recorded finite production diagnostics.

To regenerate summaries and Table 8 from the packaged per-case CSV, place the package files in the project locations expected by the runner and execute:

addpath('paper/scripts/analysis')
run_sensor_ablation('Mode', 'summarize')

Recomputing individual cases requires the complete project code, configuration, cached Kalman inputs, and validation labels. The packaged CSV files are sufficient to audit the reported event-level and pooled statistics without the raw satellite archive.