AI and Deep Learning in Hydrology and Earth Sciences
AI & Deep Learning ApplicationsDeep learning (AI) is advancing rapidly and transforming research across disciplines. While we actively adapt methods and theories pioneered in computer science to hydrology and the earth sciences, applying them reliably to real-world natural phenomena still poses significant challenges. We address these challenges by developing models and frameworks specifically tailored to hydrological and earth science applications.
1. Applying Deep Learning to Hydrological Processes
Building on rapid advances in computer science and machine learning, we apply state-of-the-art deep learning techniques to hydrological and earth science variables—including river discharge, precipitation, groundwater levels, and sea level—to improve estimation and forecasting accuracy.
River Discharge and Groundwater Level Estimation and Forecasting
We apply time-series deep learning architectures—such as Long Short-Term Memory (LSTM) recurrent neural networks—along with ensemble learning to estimate and forecast river discharge and groundwater levels, as well as to reconstruct missing observational records. (Yokoo et al., 2021; Sakaguchi et al., 2021; Nagasato et al., 2021; Yokoo, 2021a, 2021b)
Coastal Sea Level Estimation
We apply LSTM networks to model coastal sea levels. By incorporating astronomical forcing variables—such as the relative positions of the sun and moon with respect to a target site—our models accurately capture tidal variations alongside meteorological influences. (Ishida et al., 2020)
Precipitation Estimation and Downscaling
Using Convolutional Neural Networks (CNNs)—originally developed for image processing and well suited to two-dimensional spatial fields—we estimate surface precipitation from atmospheric variables. (Nagasato et al., 2020; Nagasato et al., 2021)
Hydraulic Conductivity Estimation
Using nationwide geotechnical and soil laboratory data from across Japan, we estimate hydraulic conductivity—a key physical parameter governing how readily water moves through subsurface formations—to support groundwater flow analysis.
Coastal sea level estimation using LSTM (Ishida et al., 2020)
Estimation of hydraulic conductivity from soil data
2. Evaluating the Applicability and Physical Limitations of Deep Learning
Although deep learning achieves high predictive accuracy in hydrology and the earth sciences, standard data-driven models learn statistical correlations rather than underlying physical laws (even as physics-informed approaches continue to emerge). Clarifying both the capabilities and the limitations of these models is essential before deploying them in real-world decision-making.
Input Sensitivity Analysis and Physical Interpretability
Through sensitivity analyses of input variables, we examine how individual predictors enhance or degrade model performance. In river discharge modeling with meteorological inputs, for example, we found that adding variables with a clear physical link to runoff does not necessarily improve—and can even reduce—predictive accuracy. (Nagasato et al., 2020)
Physical Consistency and Limitations of Deep Learning
A fundamental question in hydrological AI is whether data-driven models genuinely internalize physical laws from observational records. By feeding synthetic test inputs into models trained on real-world observations, we demonstrated that even highly accurate models on historical test sets can produce physically inconsistent responses outside familiar regimes. (Yokoo et al., 2021)
Evaluating physical consistency in deep learning models (Yokoo et al., 2021)
3. Developing Novel Architectures for Hydrology and Earth Sciences
Rather than simply adopting off-the-shelf architectures, we design novel deep learning frameworks and training strategies tailored to the physical and spatiotemporal characteristics of hydrological and earth science systems.
Hourly Rainfall–Runoff Modeling in Snow-Dominated Basins
In snow-affected watersheds, hourly rainfall–runoff modeling requires capturing memory over an entire seasonal cycle—an input sequence of 8,760 hourly steps per year—which standard LSTMs struggle to learn effectively. To overcome this bottleneck, we developed multi-timescale input strategies that pair recent hourly series with longer-term daily series without enforcing uniform time-step alignment (Ishida et al., 2021), as well as hybrid CNN–LSTM architectures that compress long input sequences (Ishida et al., 2024), achieving both higher accuracy and substantially faster training.
Super-Resolution of Sea Surface Temperature
We apply deep learning super-resolution techniques to enhance the spatial resolution of sea surface temperature (SST) datasets. Unlike natural photographs in computer vision, coarse- and fine-resolution geophysical datasets often differ in the physical processes they resolve, meaning off-the-shelf super-resolution methods fall short. We therefore develop domain-specific architectures and evaluation metrics tailored to SST fields. (Izumi et al., 2022)
Hybrid Physics–AI Downscaling Models
We couple physically based regional climate models with deep learning (CNNs) for hybrid precipitation downscaling, generating high-resolution spatial rainfall fields at a fraction of the computational cost required by purely dynamical simulations. (Tu et al., 2021)
Hybrid CNN–LSTM architecture for hourly rainfall–runoff modeling (Ishida et al., 2024)