Multi-step rollout forecast

When future driver channels are known (e.g. climate forecasts, calendar-based covariates, planned irrigation), a ConvLSTM trained with return_sequence = TRUE can simply be re-applied to a longer sequence covering past + future time steps — the hidden state propagates the soil memory, and every step gets its own prediction. This function automates that call and returns only the future part of the prediction.

Usage

temporal_convlstm_rollout(object, past_sequence, future_drivers)

Arguments

object

A edaphos_temporal_convlstm trained with return_sequence = TRUE.

past_sequence

Array (batch, T_past, C, H, W) — the observed window used for state warm-up.

future_drivers

Array (batch, T_future, C, H, W) with the same channel layout as past_sequence.

Value

Array (batch, T_future, H, W) with per-step predictions.