ABSTRACT

We propose a deep learning model: long short-term memory (LSTM) networks to spatially downscale Global Recovery and Climate Experiment (GRACE)-derived terrestrial water storage anomalies (TWSA) with an objective to map groundwater level anomalies (GWLA) at 0.25° resolution for basin-scale applications. Monthly TWSA from global spherical harmonic (GSH) and global mascons (GM) during 2002 to 2017 were obtained at 1° scales for the Krishna River. Eleven hydro-climatic variables were considered to observe their dependence on TWSA and further reduced to three principal components. The LSTM’s recurrent neural networks, with a 12-month lag to control flow of information in the memory units, were applied to downscale TWSA. At basin scale, downscaled GWLA from the two GRACE solutions have reasonably captured the observed trends (r > 0.6); however, GSH has underestimated the peaks (BIAS = 7.83 cm). The strong signal amplitude resulting from reduced leakage made GM a better choice over GSH in downscaling TWSA, particularly for the land–ocean mixed pixels (r_GM = 0.74, r_GSH = 0.62).

摘要

我们提出了一个深度学习模型：长短期记忆 (LSTM) 网络以在空间上缩小全球恢复和气候实验 (GRACE) 衍生的陆地蓄水异常 (TWSA)，目标是绘制 0.25° 的地下水位异常 (GWLA)盆地规模应用的分辨率。2002 年至 2017 年期间全球球谐函数 (GSH) 和全球 mascons (GM) 的月度 TWSA 是在 1° 尺度下获得的克里希纳河的。考虑了 11 个水文气候变量来观察它们对 TWSA 的依赖性，并进一步减少到三个主要成分。LSTM 的循环神经网络在控制记忆单元中的信息流方面有 12 个月的滞后，被应用于缩小 TWSA。在流域尺度上，两种 GRACE 解决方案的缩小 GWLA 合理地捕捉到了观察到的趋势（r > 0.6）；然而，GSH 低估了峰值 (BIAS = 7.83 cm)。减少泄漏导致的强信号幅度使得 GM 在缩小 TWSA 时比 GSH 更好，特别是对于陆地 - 海洋混合像素（r_GM  = 0.74，r _GSH  = 0.62）。