Skip to contents

Cross-validation folds for gauges on one stream network

Source: R/wet_cv_folds.R
wet_cv_folds.Rd

Assigns each station a fold for blocked cross-validation and describes how much a held-out station can still "see" of itself through the stations left in training, which is what makes plain leave-one-out optimistic on nested gauges.

Usage

wet_cv_folds(stations, block = c("subsubdrainage", "station"))

Arguments

stations

data.frame(station_number, wscode, localcode, upstream_area_km2), one row per station (e.g. accepted rows of wet_station_snap()).

block

"subsubdrainage" or "station".

Value

stations[, "station_number"] with fold, nesting, leak_up, leak_down.

Details

  • fold: the block, by default the WSC sub-sub-drainage (the first four characters of the station number, e.g. 08MF). All stations in a block are held out together. block = "station" gives plain leave-one-out.

  • nesting: "headwater" when no other station is upstream on the FWA network, "nested" otherwise.

  • leak_up: the largest share of the station's upstream area that a station in another fold gauges upstream of it (0 when none), so the training set already knows that much of the held-out basin.

  • leak_down: TRUE when a station in another fold is downstream, so its training residual includes the held-out basin.

Upstream is fwapg's FWA_Upstream() between the stations' fundamental watersheds (codes compared in C-locale order, as FWA codes require).