This research presents a prediction framework to simulate the net ecosystem exchange (NEE) based on a multi-perspective input selection strategy, which is an initial attempt in the pertinent literature. To accomplish the overarching aim of the current study, the data obtained by an eddy-covariance system established for monitoring three growing seasons of winter wheat were utilized. In this sense, four different input selection strategies, i.e., crop variables-based (leaf area index and total dry biomass), meteorological variables-based (solar radiation, average air temperature, soil temperature, and average relative humidity), the combination of crop and meteorological variables-based, and sensitivity analysis-based, were considered. In the last scenario, a total of 18 environmental determinants were taken into account and the most significant variables were determined based on a statistical manner via step-wise regression (SWR). The multivariate adaptive regression splines (MARS) algorithm was employed to perform the predictions, and the proposed framework was benchmarked with the artificial neural networks (ANN) as it is one of the widely used machine learning algorithms. The results revealed that the MARS model outperformed the ANN model in all scenarios but the meteorological-based. The best model performance was attained through the SWR-based MARS estimations (R-2 = 0.90 and NSE = 0.90), followed by the combination scenario (R-2 = 0.83 and NSE = 0.82), meteorological (R-ANN(2) = 0.74 and NSEANN = 0.73), and crop variable-based (R-2 = 0.72 and NSE = 0.71) scenario. Thus, this study revealed that using limited variables instead of those obtained through exhaustive measurements could yield promising performance in modelling the ecosystem fluxes, that of which can be a rewarding alternative for data-scarce regions.