Ensemble machine learning for modeling greenhouse gas emissions at different time scales from irrigated paddy fields
Ensemble machine learning for modeling greenhouse gas emissions at different time scales from irrigated paddy fields
复制标题
DOI：
10.1016/j.fcr.2023.108821
复制
发表时间：
2023-03
影响因子：
5.8
通讯作者：
Zewei Jiang;Shihong Yang;Philip Smith;Qingqing Pang
中科院分区：
文献类型：
--
作者：
Zewei Jiang;Shihong Yang;Philip Smith;Qingqing Pang
量化灌溉稻田温室气体排放对应对气候变化具有重要意义。机器学习（ML）提供了一种替代经验模型和土壤地球化学模型，但它还没有被用来模拟稻田的温室气体排放。本研究收集了中国昆山地区控制灌溉（CI，一种节水灌溉（WSI）技术）和漫灌（FI）条件下稻田的3年田间试验数据集，以及全球范围内的节水灌溉和漫灌数据集。然后，基于三个基本ML模型，随机森林（RF），K-最近邻回归（KNN），梯度提升回归（GBR）和元学习器线性回归（LR）开发了堆叠集成模型。并首次利用这些模型对WSI和FI下稻田CH 4和N2 O排放进行了不同时间尺度的模拟。结果表明，与常规稻田相比，节水灌溉稻田CH 4和N2 O的累积排放量分别减少了61.14%和19.52%。ML算法可以应用于模拟稻田温室气体日排放量、生长阶段排放量和累积排放量，但线性回归模型的模拟效果不如其他ML模型。与基本模型相比，叠加模型提高了精度，R2提高了0.37 13.36%，RMSE降低了3.23 42.78%。土壤氧化还原、温度和湿度是精确建模所必需的。同时，分别使用WSI和FI的数据对模型进行训练，有利于提高模型的精度。根据文献资料验证了叠加模型在其他台站的可用性，R2在0.7634和0.9985之间变化。因此，本文建议采用叠加模型来预测不同时间尺度下稻田的温室气体排放量，并将该方法推广应用于全球不同观测站的稻田温室气体排放量估算。
Quantifying greenhouse gas (GHG) emissions from irrigated paddy fields is of great significance for addressing climate change. Machine learning (ML) provides an alternative to empirical models and biogeochemical models, but it has not been used to simulate GHG emissions from paddy fields. In this study, a three-year field experiment dataset of paddy fields under controlled irrigation (CI, a kind of water-saving irrigation (WSI) technology) and flood irrigation (FI) in Kunshan, China, was collected, as well as a global dataset on WSI and FI. Then, the stacking ensemble model was developed based on three basic ML models, random forest (RF), K-Nearest Neighbor regression (KNN), gradient boosting regression (GBR), and a meta-learner, linear Regression (LR). Those models were used for the first time to simulate CH4and N2O emissions on different time scales from paddy fields under WSI and FI. The results showed that the cumulative CH4and N2O emissions from the WSI paddy fields decreased by 61.14% and increased by 19.52%, respectively, compared with FI. The ML algorithm can be applied to simulate daily, growth stages, and cumulative GHG emissions from paddy fields, but the performance of the linear regression model was worse than other ML models. Compared with the basic models, the stacking model improved the accuracy, improving theR2by 0.37∼13.36% and reducing theRMSEby 3.23∼42.78%. Soil redox, temperature, and moisture are necessary for accurate modeling. Meanwhile, training the model using data from WSI and FI separately is beneficial to improve the accuracy of the model. The availability of the stacking model in other stations was verified based on literature data, withR2varying between 0.7634 and 0.9985. Therefore, the stacking model is recommended to predict GHG emissions at different time scales from paddy fields, and this method can be extended to estimate GHG from paddy fields at various stations around the world.