Suspended sediment concentration estimation in the Sacramento‐San Joaquin Delta of California using long short‐term memory networks

Sedimentation is an important aspect of water resources management with many implications. Often, process‐based methods are employed to predict and assess the amount of sediment in water, but there are still challenges because the mechanisms that govern sediment transport are not yet fully understood. Furthermore, complex domains make model calibration difficult. Thus, as a complementary tool, a machine‐learning model was developed in the present study to emulate an existing process‐based model in simulating suspended sediment concentration (SSC). It employs the long short‐term memory (LSTM) networks, which are a type of artificial neural networks (ANNs) designed for supervised learning of a sequence of data (e.g., time series). The model was applied to the Sacramento‐San Joaquin Delta (the Delta) of California, USA, which is characterized by an interconnected system of sloughs, waterways, and a tidal outlet. The model training was performed with historical records of flow, stage and SSC at various locations within the Delta. The study period was 2010 through 2016, but the training period (i.e., range of observed data used to train the model) was varied to assess the model's sensitivity to the inputs and to determine the optimum model setup. Comparison between the model‐estimated SSC and the observation at 12 key locations within the Delta showed that the estimation accuracy of the LSTM model during the study period is comparable or superior to that of the Delta Simulation Model II‐General Transport Model (DSM2‐GTM), a process‐based operational hydrodynamics and water quality model for the Delta. In terms of the ratio of the root‐mean‐square error to the standard deviation (RSR), LSTM models generally showed higher predictability than DSM2‐GTM in all test cases investigated, with the lowest (most desirable) and highest (least desirable) LSTM‐based RSR being 0.21 and 1.14, respectively. In comparison, the lowest and highest RSR values with DSM2‐GTM were 0.26 and 3.70, respectively. The median LSTM‐based RSR of all study locations is around 0.7 while its DSM2‐GTM counterpart is about 1.0. LSTM models also yielded remarkably higher (more desirable) Nash‐Sutcliffe Efficiency values. Moreover, visual inspection found that LSTM models better captured the timing and magnitude of peaks as well as the temporal variations in the SSC time series. The LSTM model's performance was further analysed with hydro‐meteorological data (precipitation and wind speed) incorporat