Fit the gauge-residual adjustment of the annual water balance

Chapman et al. (2018) adjust the raw water balance (P - AET) with a regression of the gauge residuals within hydrologic zones. Here the residual obs - raw (mm) is regressed, with no global intercept, on each zone's upstream share z_k and the upstream mean of zone-masked precipitation zp_k, so each zone gets its own intercept and slope on P.

Usage

wet_wb_fit(
  d,
  min_gauges = 8,
  pooled = NULL,
  pooled_adjust = c("other", "none")
)

Arguments

d

data.frame with obs and raw (mm) and, per zone k, columns z<k> (upstream zone share) and zp<k> (upstream mean of zone-masked annual P), e.g. z08, zp08.

min_gauges

Minimum stations per zone before it is fitted on its own.

pooled

Optional character vector of zones to pool, overriding min_gauges.

pooled_adjust

"other" (one fitted level for the pooled zones) or "none" (no adjustment for them).

Value

A wet_wb_fit object: list(coef, pooled, pooled_adjust, n), where coef is a data.frame(zone, a, b) for every zone in d. Pooled zones repeat the "other" coefficients, or are 0 under pooled_adjust = "none".

Details

Every predictor is an upstream area-weighted mean of a cell layer on one analysis mask (see wet_ws_sample() and wet_upstream_means()). The fitted adjustment is therefore linear in cell values: applying it to the upstream means of any watershed (wet_wb_adjust()) gives exactly what applying it cell by cell and then averaging upstream would, and a basin that straddles zones gets the same mix of coefficients in the fit and in the application.

Zones where fewer than min_gauges stations have their largest share are pooled, decided from the stations passed in (so a cross-validation fold decides its own pooling from its training stations) unless pooled fixes it. With pooled_adjust = "other" the pooled zones share one fitted "other" level; with "none" they get no adjustment. A fitted level whose columns carry no information in the data gets coefficient 0.