Streams and rivers form dense networks that drain the terrestrial landscape and are relevant for biodiversity dynamics, ecosystem functioning, and transport and transformation of carbon. Yet, resolving in both space and time gross primary production (GPP), ecosystem respiration (ER) and net ecosystem production (NEP) at the scale of entire stream networks has been elusive so far. Here, combining Random Forest (RF) with time series of sensor data in 12 reach sites, we predicted annual regimes of GPP, ER, and NEP in 292 individual stream reaches and disclosed properties emerging from the network they form. We further predicted available light and thermal regimes for the entire network and expanded the library of stream metabolism predictors. We found that the annual network-scale metabolism was heterotrophic yet with a clear peak of autotrophy in spring. In agreement with the River Continuum Concept, small headwaters and larger downstream reaches contributed 16% and 60%, respectively, to the annual network-scale GPP. Our results suggest that ER rather than GPP drives the metabolic stability at the network scale, which is likely attributable to the buffering function of the streambed for ER, while GPP is more susceptible to flow-induced disturbance and fluctuations in light availability. Furthermore, we found large terrestrial subsidies fueling ER, pointing to an unexpectedly high network-scale level of heterotrophy, otherwise masked by simply considering reach-scale NEP estimations. Our machine learning approach sheds new light on the spatiotemporal dynamics of ecosystem metabolism at the network scale, which is a prerequisite to integrate aquatic and terrestrial carbon cycling at relevant scales.

溪流和河流形成密集的网络，排干陆地景观，并与生物多样性动态、生态系统功能以及碳的运输和转化有关。然而，到目前为止，在整个河流网络的规模上解决空间和时间的总初级生产（GPP）、生态系统呼吸（ER）和净生态系统生产（NEP）一直难以捉摸。在这里，将随机森林 (RF) 与 12 个到达站点中的传感器数据时间序列相结合，我们预测了 292 个单独的流到达中 GPP、ER 和 NEP 的年度状态，并披露了它们形成的网络中出现的属性。我们进一步预测了整个网络的可用光和热状态，并扩展了流代谢预测因子库。我们发现每年的网络规模代谢是异养的，但在春季有明显的自养高峰。与河流连续体概念一致，小的源头和较大的下游河段分别为年度管网规模的 GPP 贡献了 16% 和 60%。我们的研究结果表明，ER 而不是 GPP 在网络尺度上驱动代谢稳定性，这可能归因于 ER 的河床的缓冲功能，而 GPP 更容易受到流动引起的干扰和光可用性波动的影响。此外，我们发现大量的陆地补贴助长了 ER，这表明网络规模的异养水平出乎意料地高，否则仅仅考虑范围规模的 NEP 估计就被掩盖了。