Detecting and Validating Ecosystem Temporal Memory in Photosynthetic and Respiratory Carbon Fluxes Using Explainable Deep Learning

Abstract

Long short‐term memory (LSTM) networks effectively capture temporal dependencies in ecosystem carbon flux prediction, but whether the temporal patterns they learn reflect genuine ecological memory or arise from architectural biases remains largely unexplored. This study develops an LSTM‐attention model for predicting gross primary production (GPP) and ecosystem respiration (Reco), the two component fluxes of net ecosystem exchange (NEE), and introduces a validation procedure for integrated gradients (IG) attributions consisting of two complementary experiments. Temporal shuffling tests whether attribution patterns depend on the sequential order of the inputs, and timestep ablation confirms functional dependence on individual input months. Applied to 43 flux tower sites covering deciduous broadleaf forests (DBF), evergreen needleleaf forests (ENF), and grasslands (GRA) across North America, the models achieved test‐set R ² of 0.75–0.90 across six configurations with high attribution stability across a 10‐member ensemble. GPP showed recency‐dominant temporal memory across all three biomes, while Reco showed biome‐dependent structure, with long‐term memory in DBF ( R ² drop of 47.68% at t  = −6), recency dominance in GRA ( R ² drop of 125.26% at t  = −1), and intermediate behavior in ENF. Temporal shuffling collapsed or reorganized the ecosystem‐specific patterns, and the ablation R ² drops qualitatively corresponded to the IG attribution profiles in all six configurations. The decomposition further reveals that long‐term components of NEE memory arise predominantly from Reco and short‐term components from GPP. This work provides a transferable methodology for distinguishing ecologically meaningful temporal signals from statistical artifacts in deep learning attribution analysis, with broad applicability across the geosciences.