Machine Learning-Based Prediction of Gross Primary Productivity in Moist Deciduous Sal Forests of Western Himalayan Foothills of India
摘要
The measurement and modelling of Gross Primary Production (GPP) is critical for understanding the forest carbon and its exchange, regulating the global carbon cycle and climate. Traditionally, GPP has been measured through field-based observations or using remote sensing data, such as satellite imagery. Estimating GPP on a large scale remains difficult due to spatial and temporal variability, the complexity of forest ecosystems, and a lack of ground-based data in many regions. With advances in machine learning (ML), researchers have started leveraging these techniques to improve predictions and understanding of GPP in forest ecosystems. In this study, the performance of various ML models (viz. Artificial neural networks (ANN), Random forest (RF), Support vector machines (SVM)) and multilinear regression, for the estimation of forest GPP, is evaluated. The results indicate that both SVM (radial kernel) and RF models exhibit superior performance, achieving the lowest Root mean square error (RMSE) of 1.26 gC m−2 day−1 and 1.27 gC m−2 day−1, respectively with similar high R2 value (0.78) and Kling–Gupta efficiency (KGE) value greater than 0.80. ANN and SVM (linear kernel) performed similar with the multilinear regression with near RMSE of 1.46 gC m−2 day−1. The performance of SVM (Sigmoid kernel) was extremely poor. Among the predictor variables, air temperature has the maximum influence on the simulated GPP. In conclusion, the study suggests that the ML approach for GPP estimation is effective. However, the wise selection of the correct ML techniques, which can mimic and reproduce the complex process of forest ecosystem functioning need to be tested. The findings contribute valuable insights for researchers and practitioners seeking accurate and reliable modelling of GPP for carbon studies and environmental monitoring.