图神经网络预报印度卡纳塔克邦土壤水分：动态模型均方根误差0.022、决定系数0.919，《科学报告》11月14日刊出
《科学报告》11月14日刊出一项土壤水分预报研究。研究团队以印度卡纳塔克邦为对象，使用 2015 年 4 月至 2021 年 12 月的多源数据，提出用图神经网络预报土壤水分的时空模型。
模型设有静态与动态两种时间型土壤水分图神经网络变体，把图卷积网络、图注意力网络以及图卷积与注意力结合的架构，与长短期记忆网络结合，用于时空预报。研究将所提模型与文献中的 6 个基线模型作了比较。
结果显示，基于图注意力网络的变体整体表现最好。其动态变体的均方根误差为 0.022±0.000 立方米每立方米，决定系数为 0.919±0.003。
与表现最强的基线模型卷积长短期记忆网络相比，该动态变体的均方根误差降低了 4.34%。
论文称，结果显示图结构的时空建模在区域尺度土壤水分预报上有潜力。土壤水分预报关系到水资源管理与农业可持续性，其难点在于土壤水分随空间与时间变化大、且对环境与气候条件敏感。
图中是田野中的无线传感网络节点与灌溉管道
Graph Neural Network Forecasts Soil Moisture in Karnataka, India With Dynamic Model RMSE of 0.022 and R-Squared of 0.919, Scientific Reports, November 14
Scientific Reports published a soil moisture forecasting study on November 14. Working on Karnataka, India, the team used multi-source data from April 2015 to December 2021 and presented a graph neural network approach to forecasting soil moisture.
The researchers developed static and dynamic temporal soil moisture graph neural network variants, integrating graph convolutional networks, graph attention networks, and combined graph convolutional and attention architectures with long short-term memory networks for spatiotemporal forecasting. The proposed models were evaluated against six baseline models from the literature.
Among the evaluated models, the graph attention network based variant achieved the best overall performance. Its dynamic variant attained a root mean square error of 0.022 plus or minus 0.000 cubic metres per cubic metre and an R-squared of 0.919 plus or minus 0.003.
That corresponds to a 4.34 percent reduction in root mean square error compared with the strongest baseline, the convolutional long short-term memory network.
The paper said the results highlight the potential of graph-based spatiotemporal modelling for regional-scale soil moisture forecasting. Forecasting soil moisture matters for water-resource management and agricultural sustainability, and is challenging because of its spatial and temporal variability and its sensitivity to environmental and climatic conditions.
来源：《科学报告》
发布：GSM·美国（发布链接）
图片来源：图：田野中的无线传感网络节点与灌溉管道。作者：Stephan Brosnan，CSIRO，CC BY 3.0，维基共享资源。 / Photo: A wireless sensor network node and irrigation pipe in a field. Credit: Stephan Brosnan, CSIRO, CC BY 3.0, via Wikimedia Commons.
GSM Global · 本篇英文版 English edition of this story
