Stream water-use is essential for both agricultural and hydrological management and yet not many studies have explored its non-stationarity and nonlinearity with meteorological variables. This study proposed a deep-learning based model to estimate agricultural water withdrawal using hydro-meteorological variables, which projected the changes of agricultural water withdrawal influenced by climate change of future. The relationships between meteorological variables and stream water-use rate (WUR) were quantified using a deep belief network (DBN). The influences of precipitation, potential evapotranspiration, and monthly averaged WUR on the performance of the developed DBN model were tested. As a result, this DBN with potential evapotranspiration (PET) provided better performances than precipitation to estimate the WUR. The PET of multi-model scenarios for Representative Concentration Pathways 8.5 would be increased as time goes by, and thus leads to increase WUR estimated by DBN in three basins, located in South Korea during the future period. On the contrary, water availability expected to decrease compared to the current. Therefore, managing water-uses and improving efficiencies can be prepared for the change in agricultural water-use by climate change in the future.

中文翻译：

---

基于深度学习的 RCP 8.5 灌溉用水变化预测

河流用水对于农业和水文管理都是必不可少的，但很少有研究探讨其与气象变量的非平稳性和非线性。本研究提出了一种基于深度学习的模型，利用水文气象变量估算农业取水量，预测未来气候变化对农业取水量的影响。气象变量与河流用水率 (WUR) 之间的关系使用深度信念网络 (DBN) 进行量化。测试了降水、潜在蒸散量和月平均 WUR 对开发的 DBN 模型性能的影响。因此，这种具有潜在蒸散量 (PET) 的 DBN 提供了比降水更好的性能来估计 WUR。代表浓度路径 8.5 的多模型情景的 PET 将随着时间的推移而增加，从而导致 DBN 在未来一段时间内在位于韩国的三个盆地中估计的 WUR 增加。相反，与当前相比，可用水量预计会减少。因此，管理用水和提高效率可以为未来气候变化引起的农业用水变化做好准备。