A novel approach for retrieving GPP of evergreen forest regions of India using random forest regression

Abstract

Gross Primary Productivity (GPP) is a crucial variable of global carbon cycle for determining the ecosystem’s health. Various methods are devised to quantify GPP and upscale it in both time and space. The most common methods are physical model and eddy covariance-based estimation, which are very restricted to surrounding area of study, only. The alternative methods are empirical (e.g., LUE, CASA, SCARF, and MODIS) and Machine learning (ML) models that employ remote sensing satellite data and geographical factors. However, for using ML models, ground-based measurements of GPP is a very important factor, which is not available in most places. We propose an alternative and effective way of estimating the GPP using the ML model and data from various flux sites around the globe for a particular plant functional type (PFT). In the present study, RF is used as ML model, which is trained on global GPP data from evergreen forest and implemented in Indian region. The key findings indicated that ML-based GPP is highly accurate and hence, we generated 20 years of time series GPP dataset (2001–2020). We validated with ground-based flux tower observations during 2016–2018 for three sites (very limited datasets) in India and compared them with MODIS GPP. The coefficient of determination (R2) value of the ML-based model was 0.84 with root mean square error (RMSE) of 1.45 gC m−2 Day−1 and mean absolute error (MAE) of 0.838 gC m−2 Day−1. The proposed approach is highly accurate and far better than the MODIS-based GPP. Therefore, it can be further extended to other forest types for a holistic assessment of the carbon cycle of a region.