Decoding Carbon Flux Variability in Response to Climate Extremes in a Tropical Dry Forest Using Machine Learning

Terrestrial carbon flux dynamics are strongly influenced by climate variability, particularly in tropical dry forests (TDFs), which are drought-adapted ecosystems characterized by pronounced seasonality. However, the impacts of extreme climate events on carbon flux in TDFs remain poorly understood. In this study, we investigate the sensitivity of carbon flux to climate extremes and essential climate variables (ECVs) by integrating 11 years of net ecosystem exchange (NEE) data with 17 extreme climate indices and 8 ECVs, using a Random Forest model interpreted through Shapley Additive exPlanations (SHAP). We find that carbon flux dynamics in the SRNP-EMSS are governed primarily by ECVs rather than short-term extreme events. Results indicate that soil temperature (27.1% importance; SHAP = 0.427), vapor pressure deficit (VPD) (19.5%; SHAP = 0.265), soil moisture (13.6%; SHAP = 0.235), and air temperature (17.8%; SHAP = 0.134) emerged as the dominant drivers of carbon flux variability. These variables exhibited nonlinear responses and clear ecological thresholds: soil temperature above 26.6 °C, VPD exceeding 11 kPa, and soil moisture below 25% triggered a shift from carbon sink to source. Seasonal patterns also revealed higher carbon sourcing during dry seasons (mean SHAP = + 0.038) and greater sinking during wet seasons (mean SHAP = –0.046). While TDFs appear resilient to short-term extremes, our results highlight increasing vulnerability to sustained climatic shifts, underscoring the importance of monitoring key climate thresholds to preserve the carbon sink capacity of these ecosystems.