Prioritizing Ecosystem-Specific Carbon Exchange Drivers in Central Himalayan Forest and Grassland Using Bidirectional LSTM and SHAP Analysis

Accurate prediction of carbon exchange in the Central Indian Himalaya, a global biodiversity hotspot with extreme vertical gradients and monsoon variability, remains critical for regional carbon assessments. Despite substantial sequestration potential in Himalayan pine forests, mechanistic drivers of net ecosystem exchange (NEE) are poorly constrained, while alpine grassland carbon dynamics remain enigmatic due to observational scarcity at high elevations. Here, we integrate bidirectional long short-term memory (BiLSTM) networks with SHAP explainability to predict hourly NEE and quantitatively rank environmental controls across contrasting ecosystems. Continuous eddy covariance data from a needleleaf forest (Kosi-Katarmal, 1217 m; April-October 2020-2022) and alpine grassland (Darma Valley, 3240 m; July-October 2022-2023) were analyzed using air temperature, relative humidity, net radiation, soil conditions, vapor pressure deficit, NEE derivatives, and multi-scale lag features (1-24 h). The BiLSTM model achieved exceptional performance (forest: R² = 0.94-0.95, RMSE = 2.18-2.86 μmol m⁻² s⁻¹; grassland: R² = 0.95-0.96, RMSE = 1.09-1.73 μmol m⁻² s⁻¹). Critically, SHAP attribution unveiled fundamentally divergent control mechanisms: forest NEE was governed by rapid temporal dynamics (NEE derivative, SHAP: 0.70) and radiation-temperature coupling (SHAP: 0.02 each), signifying energy-driven photosynthetic control. Conversely, grassland NEE exhibited strong short-term memory (1-h lag, SHAP: 0.35) and atmospheric constraint dominance (temperature, SHAP: 0.06, humidity: 0.03), reflecting stomatal regulation and evaporative demand at high elevation. These findings demonstrate that forest carbon exchange operates as an energy-limited, dynamically responsive system, whereas grasslands function as atmospheric-demand limited systems with pronounced temporal persistence. Our results provide a mechanistic framework for ecosystem-specific carbon flux modeling and demonstrate the efficacy of explainable AI for process understanding in data-sparse mountain regions.