A comparison of machine learning algorithms for predicting gross primary productivity of the Western Ghats region in India using reanalysis and satellite data

Tropical forests store a substantial amount of carbon in their biomass, making them a crucial sink for atmospheric carbon dioxide (CO2). Despite a broad geographical coverage, the carbon sequestration potential of Indian forests remains an area that requires further research due to the limited availability of data and observations; this is, however, crucial for planning an efficient climate change mitigation policy development according to the Paris climate deal. The gross primary productivity (GPP) of an ecosystem is a measure of the CO2 captured by plants through photosynthesis and is considered as a key indicator of carbon sequestration. GPP is used to evaluate the carbon-capturing potential of various ecosystems. Machine learning (ML) algorithms have evolved as efficient tools for data-driven simulation in present times. In this study, we predict the GPP of one of the major forested regions in India, namely the Western Ghats, and seek to evaluate the efficacy of several ML models in this process using meteorological and biophysical variables. Our findings suggest that ML algorithms offer an alternative and efficient means for predicting GPP from meteorological and biophysical parameters. It is observed that satellite-derived leaf area index followed by vapour pressure deficit, air temperature, and soil moisture variables can be considered strong predictors for plant productivity. Our comparative analysis of multiple ML algorithms shows the decision tree regressor to perform best with an R2 score of 0.86 and mean absolute error of 3.887 gCm-2d-1. GPP plays a vital role in the terrestrial carbon cycle, and improving its prediction accuracy is beneficial for climate scientists, resource managers, and policymakers.