This study presents a new approach to understand the causes of groundwater drought events with interpretable deep learning (DL) models. As prerequisites, accurate long short‐term memory (LSTM) models for simulating groundwater are built for 16 regions representing three types of spatial scales in the southeastern United States, and standardized groundwater index is applied to identify 233 groundwater drought events. Two interpretation methods, expected gradients (EG) and additive decomposition (AD), are adopted to decipher the DL‐captured patterns and inner workings of LSTM networks. The EG results show that: (a) temperature‐related features were the primary drivers of large‐scale groundwater droughts, with their importance increasing from 56.1% to 63.1% as the drought events approached from 6 months to 15 days. Conversely, precipitation‐related features were found to be the dominant factors in the formation of groundwater drought in small‐scale catchments, with the overall importance ranging from 59.8% to 53.3%; (b) Seasonal variations in the importance of temperature‐related factors are inversely related between large and small spatial scales, being more significant in summer for larger regions and in winter for catchments; and (c) temperature‐related factors exhibited an overall “trigger effect” on causing groundwater drought events in the studying areas. The AD method unveiled how the LSTM network behaved differently in retaining and discarding information when emulating different groundwater droughts. In summary, this study provides a new perspective for the causes of groundwater drought events and highlights the potential and prospect of interpretable DL in enhancing our understanding of hydrological processes.
Soil moisture‐Precipitation (SM‐P) coupling, especially the causality from SM to P (SM → P), is an important and highly debated topic. The causal inference from observational data provides a statistical approach for this issue, where the experimental research is infeasible. Various causal inference methods exist, each assuming distinct underlying systems for the targeted variables: pure stochastic or deterministic dynamic system (DDS), which means that these methods detect different types of causality: separable or non‐separable. Acknowledging the inherent deterministic dynamic nature of SM‐P coupling, this study employs a DDS‐based method: Convergent Cross‐mapping (CCM), to detect their non‐separable causality. Centering around SM‐P coupling, we also detect the causality between SMs in shallow and deeper layers, and the causality between evapotranspiration (ET) and SMs in all layers. Notably, before applying CCM, a preconditional procedure is required: verifying the DDS nature of targeted variables P, SM, and ET. Key findings of this research include: (a) In SM‐P coupling, only SMs in shallow layers but not deeper layers could render causalities toward P, while P renders causalities toward SMs in all layers. (b) The time‐delay of causality from SM1 to P in spring/summer is around 4–6 days, and that from P to SM1 is around 2–4 days. (c) Inside SMs, shallow SMs have obvious causalities downwards to deeper SMs, but deeper SMs seem harder to render causalities upwards to shallow SMs, explaining why they hardly render causalities to P. In summary, this study furnishes an indispensable DDS‐based complement to stochastic‐based methodologies for SM‐P coupling.
Vegetation, a key intermediary linking water, the atmosphere, and the ground, performs extremely important functions in nature and for our existence. Although satellite-based remote-sensing technologies have become important for monitoring vegetation dynamics, selecting the correct remote-sensing vegetation indicator has become paramount for such investigations. This study investigated the consistencies between a photosynthetic activity index (the solar-induced chlorophyll fluorescence (SIF) indicator) and the traditional vegetation index (the Normalized Difference Vegetation Index (NDVI)) among different land-cover types and in different seasons and explored the applicability of NDVI and SIF in different cases by comparing their performances in gross primary production (GPP) and grain-yield-monitoring applications. The vegetation cover and photosynthesis showed decreasing trends, which were mainly concentrated in northern Xinjiang and part of the Qinghai–Tibet Plateau; a decreasing trend was also identified in a small part of Northeast China. The correlations between NDVI and SIF were strong for all land-cover types except evergreen needleleaf forests and evergreen broadleaf forests. Compared with NDVI, SIF had some advantages when monitoring the GPP and grain yields among different land-cover types. For example, SIF could capture the effects of drought on GPP and grain yields better than NDVI. To summarize, as the temporal extent of the available SIF data is extended, SIF will certainly perform increasingly wide applications in agricultural-management research that is closely related to GPP and grain-yield monitoring.