Gross primary productivity estimation through remote sensing and machine learning techniques in the high Andean...

Accurately estimating gross primary productivity (GPP) is crucial for simulating the carbon cycle and addressing the challenges of climate change. However, estimating GPP is challenging due to the absence of direct measurements at scales larger than the leaf level. To overcome this challenge, researchers have developed indirect methods such as remote sensing and modeling approaches. This study estimated GPP in a humid páramo ecosystem in the Andean Mountains using machine learning models (ML), specifically Random Forest (RF) and Support Vector Regression (SVR), and compared them with traditional models. The study's objective was to analyze the strength and complex nonlinear relationships that govern GPP and to perform an uncertainty analysis for future climate projections. The methodology used to estimate GPP showed that ML-based models outperformed traditional models. The performance of ML models varied significantly among seasons, with the correlation coefficient (R) ranging from 0.24 to 0.86. The RF model performed better in capturing the temporal changes and magnitude of GPP in the less humid season, displaying the highest R (0.86), lowest root mean squared error (0.37 g C*m^-2), and percentage bias (-3%). Additionally, the analysis indicates that solar radiation is the primary predictor of GPP in the páramo biome, rather than water. The study presents a method for deriving daily GPP fluxes and evaluates the impact of various variables on GPP estimates. This information can be employed in the development of vegetation prediction models.