Model Drivers

Environmental Predictors Driving Spatial Variation in Downscaled Predictions

The Random Forest downscaling model learns relationships between environmental site characteristics and SIPNET model outputs at 100 anchor sites, then predicts to all ~132,000 cropland fields using eight predictors (seven environmental covariates plus field area). The diagnostic plots below reveal which environmental factors drive the spatial variation seen in the maps. See the Scenario Definition Table for practice parameters and simulation details.

Validation Status

These results are from a proof-of-concept modeling framework that has not been validated against field observations. Interpret all values as illustrative projections, not empirical estimates.

Model Performance Overview

Before interpreting predictor effects, it is important to understand how well each Random Forest model captures spatial variation. Out-of-bag (OOB) R² measures predictive accuracy on held-out training sites.

Variable importance across all four response variables. Importance is measured as the median increase in MSE when each predictor is permuted, with IQR across 20 ensemble members.

Downscaling model performance by response variable. R² is the median across 20 ensemble members.

Variable	Median OOB R²	IQR	Top Predictors	Assessment
N2O Flux	0.52	0.38–0.72	vapr, temp	Best spatial model
TotSoilCarb	0.34	0.26–0.41	ocd, vapr	Moderate
AGB	0.09	-0.04–0.52	precip, vapr	Poor – high variance across ensembles
CH4 Flux	-0.15	-0.17–-0.13	none meaningful	No spatial signal (near-zero baseline)

Interpreting CH₄ Model Quality

Non-flooded annual croplands produce near-zero CH₄ fluxes in SIPNET (approximately 10⁻¹⁴ kg m⁻² at the modeled sites) because methanogenesis requires sustained anaerobic conditions (continuous flooding) that do not occur under standard irrigation. With essentially no spatial variation in the training data, the Random Forest has no signal to learn – the negative OOB R² (-0.15) confirms the model performs worse than a constant mean. The CH₄ ALE/ICE plots below reflect noise, not ecology.

Environmental Predictors

The downscaling model uses eight predictors (seven environmental covariates plus field area):

Covariates used in the Random Forest downscaling model.

Variable	Description	Source	Units	Ecological Role
clay	Clay content	SoilGrids	%	Soil texture controls water holding capacity and physical protection of organic matter
ocd	Organic carbon density	SoilGrids	hg/m3	Existing soil organic matter – both a predictor and correlated with the output
twi	Topographic wetness index	SRTM-derived	-	Topographic position controls water accumulation and drainage
temp	Mean annual temperature	ERA5	°C	Temperature drives decomposition rates and plant productivity
precip	Mean annual precipitation	ERA5	mm/yr	Moisture availability for plant growth and microbial activity
srad	Solar radiation	ERA5	J/m²	Energy input for photosynthesis
vapr	Vapor pressure	ERA5	kPa	Atmospheric moisture – strongly correlated with temperature (r = 0.86)
area_ha	Field area	CADWR	ha	Structural covariate capturing field-size effects on model predictions

ALE and ICE Diagnostic Plots

ALE (Accumulated Local Effects) plots show the average marginal effect of each predictor on the model output across its range. Steeper slopes indicate stronger influence.

ICE (Individual Conditional Expectation) plots show how each individual anchor site responds to changes in a predictor while holding all others constant. When ICE lines diverge, it indicates interaction effects – the predictor’s influence depends on other site characteristics.

---

Total Soil Carbon (TotSoilCarb)

ALE – Top Predictors

TotSoilCarb ALE – Predictor 1

TotSoilCarb ALE – Predictor 2

Higher organic carbon density (ocd) –> higher soil carbon: sites with more existing organic matter retain more carbon through the simulation. Vapor pressure (vapr) shows a negative relationship – higher vapr (warmer, more humid conditions) correlates with lower soil carbon, consistent with faster decomposition at higher temperatures (vapr and temp are highly correlated, r = 0.86).

ICE – Top Predictors

TotSoilCarb ICE – Predictor 1

TotSoilCarb ICE – Predictor 2

ICE lines that move broadly in parallel indicate the predictor has a consistent effect across different site types. Diverging lines would indicate interactions with other environmental gradients.

---

Aboveground Biomass (AGB)

ALE – Top Predictors

AGB ALE – Predictor 1

AGB ALE – Predictor 2

Precipitation (precip) and vapor pressure (vapr) are the top drivers of AGB variation, though this model has low explanatory power (median OOB R² = 0.09). Higher precipitation –> more water availability –> more growth. Vapor pressure is strongly correlated with temperature (r = 0.86); its negative relationship with AGB likely reflects faster turnover under warmer conditions rather than a direct moisture effect. The wide IQR (-0.04 to 0.52) across ensembles indicates the spatial pattern is not reliable.

ICE – Top Predictors

AGB ICE – Predictor 1

AGB ICE – Predictor 2

---

N₂O Flux

ALE – Top Predictors

N₂O ALE – Predictor 1

N₂O ALE – Predictor 2

Vapor pressure (vapr) and mean annual temperature (temp) are the top predictors of spatial variation in N₂O flux (OOB R² = 0.52, our best spatial model). Higher vapr and temperature are associated with greater N₂O emissions, consistent with warmer conditions promoting nitrification and denitrification. These two predictors are highly correlated (r = 0.86), reflecting California’s north-south climate gradient.

ICE – Top Predictors

N₂O ICE – Predictor 1

N₂O ICE – Predictor 2

---

CH₄ Flux

ALE – Top Predictors

CH₄ ALE – Predictor 1

CH₄ ALE – Predictor 2

Low Model Quality

The CH₄ downscaling model has negative OOB R² (-0.15), indicating it has no predictive skill for spatial variation. Non-flooded annual croplands produce near-zero CH₄ fluxes (raw SIPNET output on the order of 10⁻¹⁴ kg m⁻²; statewide downscaled total roughly 13,600 kg CH₄ yr⁻¹) because methanogenesis requires sustained soil saturation that does not occur under standard irrigation, leaving the Random Forest with no spatial signal to learn. The top “predictors” (area_ha, clay) carry no ecological meaning here – they reflect noise fitting. Interpret these plots accordingly.

ICE – Top Predictors

CH₄ ICE – Predictor 1

CH₄ ICE – Predictor 2

---

Why These Drivers Matter

Understanding what drives spatial variation is essential for targeting conservation practices:

1. If soil type dominates (clay, ocd) – practice benefits are site-specific, and adoption should prioritize fields with responsive soil types.
2. If climate dominates (temp, vapr, precip) – practice effects vary with regional climate, suggesting geographic targeting based on climate zones.
3. If the practice effect is uniform regardless of drivers – broad-scale adoption is efficient because benefits are not geographically restricted.

The field-level difference maps in the Soil Carbon Atlas and GHG Emissions Atlas show which pattern applies for each variable and scenario.