Enhanced Carbon Flux Forecasting via STL Decomposition and Hybrid ARIMA-ES-LSTM Model in Amazon Forest

Abstract

This study presents a hybrid model, STL-ARIMA-ES-LSTM, developed to improve the accuracy of Gross Primary Productivity (GPP) forecasts in the Amazon region. The model integrates seasonal and trend decomposition using Loess (STL) with statistical methods (ARIMA and Exponential Smoothing-ES) and a machine learning technique (Long Short-Term Memory - LSTM). Applied to GPP data from the PE-QFR site, the hybrid model achieved significantly better error metrics, with RMSE of 1.69 gC/m2/day, MAE of 1.35 gC/m2/day, and MAPE of 0.20%, compared to the standalone LSTM (RMSE of 2.16 gC/m2/day, MAE of 1.78 gC/m2/day, and MAPE of 0.27%). Furthermore, the hybrid model showed stronger agreement with the observed data, with correlation coefficient r =0.62 and ${R}^{2} =0.39$ , whereas the LSTM alone yielded r =0.26 and ${R}^{2} =$ –0.002. The STL decomposition allowed effective separation of trend, seasonality, and residual components, enabling tailored modeling of each, which contributed to the improved predictive performance. These results demonstrate the advantage of hybrid approaches in capturing the nonlinear and seasonal patterns of GPP, supporting enhanced environmental monitoring and more informed climate change mitigation strategies in the Amazon.

Keywords

• Gross primary productivity
• carbon flux
• hybrid models
• STL decomposition
• LSTM
• statistical models