Physical Quantities

Which atmospheric variables actually carry the fuel moisture signal? Every quantity added multiplies through the history dimension — one more QoI at 8 historical times is 8 more features — so dropping an uninformative variable is worth real compute.

These studies use datasets 39–42, which differ only in spatial sample size (1,000 to 4,000 grid points at 2,000 reference times), with \(t_{max\_history} = 32\) h and \(t_{history} = 4\) h. Varying the feature set while holding the data fixed is exactly what qois_for_training is for.

Effect of elevation

Elevation (HGT) is the only feature that does not vary with time. Each dataset was trained with and without it.

Dataset	Grid points	Without HGT	With HGT	Gain
39	1,000	0.8476	0.8549	+0.0073
40	2,000	0.8701	0.8772	+0.0071
41	3,000	0.8835	0.8906	+0.0071
42	4,000	0.8933	0.9005	+0.0072

Elevation gives a small but remarkably consistent benefit — almost exactly +0.007 in R² regardless of dataset size. The stability of that number across four independent datasets is itself evidence the effect is real rather than noise.

Correlation between ground truth and predicted FM for datasets trained with and without elevation.

At one feature out of 41, elevation is cheap. Keep it.

Effect of precipitation and shortwave flux

Results not yet in the manuscript

This section is a placeholder in the manuscript; the numbers come from the simulation output.

Precipitation (PREC) and downward shortwave flux (SW) were dropped individually and together, from a base set of HGT, UMag10, T2, RH.

Features	Dataset 39	Dataset 40	Dataset 41	Dataset 42
HGT, UMag10, T2, RH, PREC, SW	0.8549	0.8772	0.8906	0.9005
HGT, UMag10, T2, RH, SW (no PREC)	0.8459	0.8695	0.8828	0.8933
HGT, UMag10, T2, RH, PREC (no SW)	0.8236	0.8412	0.8540	0.8630
HGT, UMag10, T2, RH (neither)	0.8094	0.8291	0.8399	0.8494

On dataset 42, relative to the full set:

• Dropping PREC costs 0.0072
• Dropping SW costs 0.0375 — over five times as much
• Dropping both costs 0.0511

Shortwave downward flux is far more important than precipitation. The ordering is identical on all four datasets, and the same pattern reappears independently in the maximum history study, where dropping SW cost 0.0369 and dropping PREC cost 0.0079 on a completely different set of datasets.

The physical reading is straightforward: solar flux drives the drying of fine fuels directly and operates continuously, whereas precipitation is intermittent and mostly zero in California outside the winter months. A variable that is zero in the majority of samples carries little information for most predictions, however decisive it is when it does occur.

Tip

If feature count must be reduced, PREC is the cheapest thing to drop. SW is not.

Temperature, humidity, and vapor pressure deficit

Results not yet in the manuscript

This section is a placeholder in the manuscript; the numbers come from the simulation output.

Vapor pressure deficit combines temperature and relative humidity into one variable (see Step 2). If VPD could replace both, the feature count would drop by one QoI — 8 features at the baseline history settings. Six combinations were tested against a base set of HGT, UMag10, PREC, SW.

Additional features	Dataset 39	Dataset 40	Dataset 41	Dataset 42
T2, RH (baseline)	0.8549	0.8772	0.8906	0.9005
T2, VPD	0.8499	0.8714	0.8851	0.8953
RH, VPD	0.8459	0.8670	0.8808	0.8905
RH alone	0.8384	0.8601	0.8751	0.8843
VPD alone	0.8364	0.8592	0.8740	0.8842
T2 alone	0.8265	0.8540	0.8726	0.8840

Three findings, all consistent across the four datasets:

VPD does not replace temperature and humidity. Substituting VPD for both costs 0.0163 on dataset 42 (0.9005 → 0.8842). The combination of T2 and RH outperforms every alternative tested.

VPD alone is no better than either input alone. VPD (0.8842), RH (0.8843) and T2 (0.8840) are within 0.0003 of each other on dataset 42 — an effective three-way tie. Whatever VPD gains by combining the two variables, it loses by discarding their independent information.

The best substitution keeps temperature. T2, VPD (0.8953) recovers most of the baseline and clearly beats RH, VPD (0.8905), suggesting the residual signal in T2 that VPD does not capture is more valuable than the equivalent in RH.

Is the substitution worth it?

Option	R² (dataset 42)	QoIs	Features at baseline history
T2, RH	0.9005	6	49
T2, VPD	0.8953	6	49
VPD alone	0.8842	5	41

Replacing T2, RH with VPD alone saves 8 features — about 16% — for 1.6 percentage points of R². Whether that trade is worth taking depends on the application, but the motivating hope that VPD would be a free consolidation is not supported by these runs.

Summary

Variable	Verdict
SW (shortwave flux)	Important — dropping costs ~0.037
T2 + RH together	Important — best available pairing
HGT (elevation)	Small but consistent gain, ~0.007, for one feature
PREC (precipitation)	Marginal — dropping costs ~0.007
VPD	A viable economy, not a free win

The recommended full feature set is HGT, UMag10, T2, RH, PREC, SW, which is what the highest-scoring configuration in every study on this page uses.

Not yet assessed

Terrain ruggedness appears in the manuscript outline as a candidate feature but has no simulation results. Given that elevation alone contributes a small, stable gain, a ruggedness measure derived from the elevation field is a reasonable next test.