Modeling GPP with Machine Learning using Multisource Features based on Fluxnet Data

One of the most crucial markers for forecasting the future trend of climate change and for adopting sustainable development plans is the understanding of the carbon that plants absorbs from the atmosphere. Terrestrial Gross Primary Productivity (GPP), a measure of carbon uptake, is the total amount of carbon that is assimilated by photosynthesis in a terrestrial ecosystem. Eddy covariance measurements taken from Flux-towers are regarded as accurate GPP estimates. However, building a tower is expensive and difficult to maintain, thus we only have a small number of towers for flux measurements. Because satellite data have a greater resolution and continuous coverage, we can use them create models to estimate GPP at larger geographical and temporal scales. Consequently, a method based on machine learning (ML) technique that directly affects the GPP by leveraging external elements such as remotely sensed data, geographic data, and meteorological data. We propose a random forest model that had an R
²
value of 0.82 for estimating GPP based on 10-fold cross-validation, in the Australian region, and the findings are contrasted with those obtained using current state-of-the-art techniques like SVM and XGBoost. Performance of future prediction has been demonstrated by predicting the GPP for year 2014 for which the R
²
value attained 0.84 with respect to the ground truth, whereas the MODIS GPP was only 0.52. Therefore, these characteristics can be used to model the GPP and can forecast values across a range of time scales and with varied levels of spatial detail.