MemPerceiver: Adaptive multi-scale memory reveals biome-specific temporal fingerprints in carbon flux prediction

Terrestrial carbon flux prediction is critical for quantifying ecosystem carbon sources and sinks. However, current deep learning models still struggle with cross-ecosystem generalisation across the heterogeneous biomes covered by global eddy-covariance networks. A fundamental reason is that different ecosystems are shaped by past climate over very different timescales: climate may continue to influence ecosystem processes for only weeks in tropical savannas but for months in temperate forests, while existing models apply a single uniform temporal context window to all ecosystems. To address this challenge, we propose MemPerceiver, a multimodal framework that augments the EcoPerceiver backbone with six ecologically motivated memory scales (−180 to −7 days) and a per-scale gating network, allowing the model to autonomously select which temporal horizons matter for each ecosystem without any ecological prior. On the CarbonSense benchmark of 385 eddy-covariance sites across 15 IGBP classes, MemPerceiver achieves state-of-the-art performance on both Net Ecosystem Exchange (NEE) and Gross Primary Production (GPP), with the largest gains where cross-ecosystem generalisation is hardest: a 23% RMSE reduction on the one-shot WAT class and a 24% NSE improvement on the zero-shot SNO class. Averaging the learned gate activations within each biome yields an interpretable temporal fingerprint, and within-biome consistency analysis shows that these fingerprints provide a stable, model-derived view of biome-specific ecological memory. The code is available at https://anonymous.4open.science/r/MemPerceiver-C000.