Civil and Environmental Engineering
Abstract: Deep Learning (DL) methods have made revolutionary strides in recent years. A core value proposition of DL is that abstract notions and patterns can be extracted purely from data, without the need for domain expertise. Process-based models (PBM), on the other hand, can be regarded as repositories of human knowledge or hypotheses about how systems function. Here, through computational examples, we argue that there is merit in integrating PBMs with DL due to the imbalance and lack of data in many situations. We trained a big-data time series deep learning neural network to learn and hindcast soil moisture dynamics from Soil Moisture Active Passive (SMAP) Level 3 product. As the first application of time series deep learning in hydrology, we show promising hindcasting results and that DL is more robust than simpler methods. With high fidelity to SMAP products, our data can aid various applications including data assimilation, weather forecasting, and soil moisture hindcasting. PBMs have some generalization value which should be carefully assessed and utilized. On the other end, we demonstrate our capability to simulate large-scale seasonal flood inundation using an adaptive multi-mesh multi-resolution hydrologic model. The simulation explains the seasonal water composition in a floodplain lake adjacent to the Amazon River. We envision that in the future these two domains can be coupled to produce enhanced flood inundation predictions.