Local Sensitivity Analysis

This report ranks the parameters of the SIPNET ecosystem model by their contribution to the uncertainty of modeled soil carbon and nitrous oxide flux. The analysis covers 100 design points on California cropland over 2016 to 2023. Results are each parameter’s share of parameter variance at each design point.

Background

A parameter matters for prediction when the model is sensitive to it and its value is uncertain. The analysis combines the two, following the PEcAn variance decomposition (LeBauer et al., 2013; Dietze et al., 2014).

SIPNET is run with every parameter at its prior median, then with each parameter moved in turn to the 2.3, 15.9, 84.1, and 97.7 percentiles of its prior, all others at their median. A spline through the five outputs gives the response to that parameter, and the variance of the prior propagated through the spline is its contribution to the output variance. The share of parameter variance is that contribution over the sum across all plant and soil parameters at the design point. A parameter is treated as influential where its share exceeds 5 percent, the criterion Fer et al. (2018) used to select SIPNET parameters for calibration (Fer et al., 2018).

Figure 1: Design points, colored by the PFT they were run under, with the number of design points in brackets. Outlines are the Cal-Adapt climate regions.

Table 1: Configuration of the analysis.

Setting	Value
Model	SIPNET 2.2.0, nitrogen cycle and anaerobic processes on
Design points	100: 67 woody perennial, 14 row crop, 6 grass, 4 alfalfa, 3 rice, 2 corn, 4 uncropped
Period	2016 to 2023, reported as the mean over the period
Parameters	34 per cropped design point (21 plant, 13 soil), 13 per uncropped design point
Priors	PFT priors of the statewide inventory runs
Meteorology, initial conditions, management	member 1 of each inventory ensemble
Model runs	13,364

Share of parameter variance

Figure 2: Prior uncertainty (CV), sensitivity (elasticity), and share of parameter variance. Points are medians across design points and bars span the interquartile range.

Table 2: Share of parameter variance across the 100 design points, for parameters with a median share above 5 percent.

Output	Parameter	Median share (%)	Range (%)	Design points over 5 percent
Soil carbon	Soil respiration Q10	62	4 to 78	99
	Base soil respiration rate (0 °C, saturated soil)	28	2 to 42	98
Nitrous oxide flux	Soil respiration Q10	46	17 to 81	100
	Nitrogen volatilization rate constant	40	15 to 67	100
	Nitrogen leaching fraction	8	0 to 20	76

Two soil parameters account for most of the parameter variance in soil carbon, and raising either lowers soil carbon at all 100 design points. A plant parameter exceeds 5 percent of the soil carbon variance at 19 of the 96 cropped design points, 18 of them woody perennial, and at none for nitrous oxide.

Variation across design points and PFTs

Figure 3: Share of parameter variance across design points. Upper panels: distribution across design points, with the mean marked by the green bar. Lower panels: mean share by PFT.

The two parameters with the largest mean share are the same in every PFT for each output.

Response to each parameter

Figure 4: Prior density and model response for the soil parameters at one row crop design point in the San Joaquin Valley. Points are the five runs, the larger point the median run, and the line PEcAn’s spline.

Figure 5: Response of soil carbon at every design point, as output over the design point’s median run, from the 2.3 to the 97.7 percentile of each prior.

Figure 6: Response of nitrous oxide flux at every design point, as in Figure 5.

Reporting scope

Shares are reported for soil carbon and nitrous oxide flux. Methane and six other outputs were also computed. Modeled methane is near the precision of the model output, and plant carbon is low in these runs, with net primary productivity below 20 g C m^-2 yr^-1 at the median design point of every PFT; shares for these outputs are not presented.

Interpretation

These are model-derived results. Parameters are drawn from priors that have not been updated with observations, so each share depends on the prior widths; the prior for the soil respiration Q10 is uniform from 1.4 to 5.0.

Each parameter is varied with all others at their medians, so the shares divide the sum of single-parameter contributions and do not include interactions (Saltelli et al., 2019). Meteorology, initial conditions, and management are held at one ensemble member; the global sensitivity analysis compares those inputs with the parameters. In SIPNET 2.2.0 the nitrous oxide output is the total nitrogen volatilization flux.

See References for the cited sources.