Accurately quantifying carbon fluxes across ecosystems is essential for monitoring and validating natural climate solutions (NCS), which promise to mitigate climate change. Measurement methods, such as eddy covariance towers, provide ground truth data at high temporal resolution but suffer from limited spatial coverage. Upscaling these measurements to ecosystem scales is performed with machine learning methods based on environmental drivers and satellite data. However, correctly quantifying uncertainty in these predictions remains a challenge, which limits its use in carbon markets. We propose an uncertainty-aware carbon flux estimation framework that integrates multispectral Landsat imagery, EC flux measurements, and ancillary environmental variables using Mixture Density Networks. Our framework provides estimates of both aleatoric and epistemic uncertainties that enhance the reliability and scalability of carbon monitoring efforts.