Skip to contents

Temporal Active Learning: rank candidate cells by their Kalman gain norm after the latest EnKF assimilation

Source: R/temporal_active_learning.R
al_query_temporal.Rd

Closes the loop between Pillar 3 (4D ConvLSTM + stochastic EnKF assimilation) and Pillar 5 (autonomous active learning). The Kalman-gain norm is a forward-looking estimate of how much the posterior ensemble would shrink if a new observation were placed at each spatial cell; candidates with high gain are the natural next sampling locations.

Usage

al_query_temporal(
  kalman_update,
  candidate_coords = NULL,
  n_select = 10L,
  combine = c("gain", "gain_sd", "gain_sd_normalised")
)

Arguments

kalman_update

A edaphos_temporal_kalman object returned by temporal_kalman_update(). Must carry the gain_row_norm and analysis_sd fields (they are produced by default).

candidate_coords

Optional data frame with lon, lat columns that restrict scoring to a finite set of physically accessible cells. When NULL, every cell of the analysis grid is a candidate.

n_select

Integer; how many candidates to return.

combine

One of "gain" (rank by pure gain norm, the v1.5.0 default), "gain_sd" (weighted product of gain and remaining analysis SD), or "gain_sd_normalised" (same but each term is percentile-normalised first).

Value

Data frame sorted by descending priority with columns row, col, gain, analysis_sd, priority.