13 MAY 2026 (WED) 14:35 - 15:05

Ecosystem Methane Fluxes Estimation Using Causal Spatiotemporal Machine Learning

Miss LI Mengyao   

( Supervisor: Prof Hongsheng Zhang )

Abstract:

Methane (CH4) is the second most important anthropogenic greenhouse gas, accounting for approximately 20% of total radiative forcing and roughly half of the net global warming since pre-industrial times. Global CH4 emissions originate from a diverse mosaic of ecosystem types, including tropical floodplains, boreal peatlands, temperate freshwater wetlands, rice paddies, coastal blue carbon ecosystems, and urban-industrial landscapes, each regulated by distinct combinations of biological, hydrological, and anthropogenic drivers, whose causal interactions remain poorly characterized. Existing approaches to CH4 flux estimation rely predominantly on associational machine learning models that conflate causal and spurious relationships, producing predictions accurate within observed conditions but incapable of supporting policy-relevant interventional analysis or process-level mechanistic understanding.

This study develops a causality-based progressive causal learning framework that systematically advances from causal discovery through causal machine learning and causally constrained deep learning to process-level attribution and counterfactual scenario analysis. The experiment was based on data from 32 globally distributed sites spanning 7 ecosystem land-use types and diverse urban gradients. Firstly, PCMCI+ causal discovery or the Conditional Average Treatment Effect methods were applied to the multivariate environmental time series at each site to identify ecosystem-specific causal relationships and directed acyclic graphs governing CH4 fluxes. These causal graphs are then embedded as architectural priors in the Causal Temporal Graph Attention Network (CT-GAT), which generates monthly CH4 flux data products through causally constrained temporal encoding and graph-attention-based spatial aggregation. Finally, the fitted structural causal model is applied to causal mediation analysis, counterfactual scenario evaluation, and the construction of a spatially explicit, ecosystem-stratified CH4 budget.

Preliminary results validate the framework's effectiveness. The Causal Random Forest model achieves the lowest prediction error (R2 = 0.940), and CT-GAT outperforms 11 association-based baselines (R2 = 0.945). PCMCI+ identifies atmospheric pressure, vapor pressure deficit, and air temperature as dominant positive causal drivers of CH4 flux across biome types, with relative humidity as the primary suppressive influence and land cover class as a structural hub that modifies the causal pathway architecture. Hierarchical mixed-effects analysis reveals systematic sign reversals in the effects of temperature, wind speed, and GPP along the urbanization gradient, providing quantitative evidence of a structural transition from biogenic to anthropogenic CH4 dominance in developed landscapes. Spatiotemporal mapping reveals regionally divergent decadal CH4 trends and land-use-specific seasonal signatures consistent with the identified causal mechanisms. The framework is designed to be extended to CO2 and N2O, ultimately supporting comprehensive multi-gas greenhouse gas characterization across globally diverse ecosystems.