Abstract Continuous N2O flux acquisition is crucial to enrich our knowledge of the complex mechanisms underlying the annual greenhouse gas budget and to refine their estimation. N2O flux measurement methodologies at high temporal resolution, i.e. micro-meteorology methodologies, are still scarce and may exacerbate the lack of important data, especially during the night if the required turbulent conditions are not met. The static and automated chamber methodologies also lead to numerous gaps in a time series due to low sampling frequency, hardware malfunctions, chambers removal during field operations or filtering of low-quality measurements. There is a strong need to define a generic and realistic N2O flux gap-filling methodology, especially since there is no consensus on the methodology to be used. In this study, we investigated the effect of using either the traditional linear interpolation methodology alone, either an Artificial Neural Networks (ANN) methodology alone or the combination of both on gap-filled daily N2O flux dynamics and annual budget. All three methodologies were tested on daily N2O flux time series measured with automated chambers over 5 years from 2012 to 2016 on a southwestern France crop site following a wheat – maize rotation. On average over the studied period, the results showed better statistical scores using the ANN methodology alone than using the linear interpolation methodology alone, with R² and RMSE of 0.84 and 12.4 gN ha−1 d−1 and of 0.68 and 17.4 gN ha−1 d−1, respectively. However, whereas the use of ANN methodology reproduced well high measured N2O fluxes, it induced overestimation on low measured N2O fluxes where the use of the linear interpolation methodology was relevant. To overcome that issue and to take advantages of both methodologies we propose a new one which mixes both. On average, using the mixed methodology did not increase statistical scores compared to the ANN one, with a R² and a RMSE of 0.84 and 12.4 gN ha−1 d−1 respectively for both, but for periods with low measured N2O fluxes using the mixed methodology improved the statistical scores and the observed daily flux dynamic.

中文翻译：

---

结合人工神经网络和线性插值两种方法，填补每日一氧化二氮通量测量的空白

摘要 连续获取 N2O 通量对于丰富我们对年度温室气体预算背后复杂机制的了解并完善其估算至关重要。高时间分辨率的 N2O 通量测量方法，即微气象学方法，仍然稀缺，可能会加剧重要数据的缺乏，尤其是在不满足所需湍流条件的夜间。由于低采样频率、硬件故障、现场操作期间的腔室移除或低质量测量的过滤，静态和自动化腔室方法还导致时间序列中的大量间隙。非常需要定义一个通用的和现实的 N2O 通量间隙填充方法，特别是因为对要使用的方法没有达成共识。在这项研究中，我们研究了单独使用传统线性插值方法、单独使用人工神经网络 (ANN) 方法或两者结合使用对间隙填充的每日 N2O 通量动态和年度预算的影响。所有这三种方法都在 2012 年至 2016 年法国西南部一个小麦-玉米轮作后的作物种植区使用自动化室测量的每日 N2O 通量时间序列进行了测试。平均而言，在研究期间，结果显示单独使用 ANN 方法比单独使用线性插值方法具有更好的统计分数，R² 和 RMSE 分别为 0.84 和 12.4 gN ha-1 d-1，以及 0.68 和 17.4 gN ha-1分别为 d−1。然而，虽然使用 ANN 方法可以很好地再现高测量的 N2O 通量，它导致对低测量 N2O 通量的高估，其中使用线性插值方法是相关的。为了克服这个问题并利用这两种方法，我们提出了一种将两者混合的新方法。平均而言，与 ANN 相比，使用混合方法并没有增加统计分数，两者的 R² 和 RMSE 分别为 0.84 和 12.4 gN ha-1 d-1，但对于使用混合方法测得的 N2O 通量较低的时期方法改进了统计分数和观察到的每日通量动态。