Hydrological signatures (HS) have proven to be highly effective in calibrating physically‐based hydrological models, enhancing their process consistency. However, their integration into parameter optimization for deep learning (DL)‐based hydrological models has been limited. To address this gap, we propose a novel HS‐informed framework that dynamically integrates HS into DL parameterization through a multi‐task learning approach. This study evaluates the impact of HS integration on model performance using a large‐scale, global hydrological data set. The HS‐informed model achieved a significant performance improvement, with a median Nash‐Sutcliffe Efficiency (NSE) of 0.739, compared to 0.666 for the baseline model across the test set. Notably, the most pronounced improvements in NSE were observed in hydrologically complex basins, including baseflow‐dominated (+0.135), drought‐prone (+0.148), and flood‐prone basins (+0.159). Sensitivity analysis further revealed that the HS‐informed model could leverage extended historical input data (over 120 days) to sustain robust performance (median NSE of 0.715) over a 30‐day forecast period. Shapley Additive Explanations analysis highlighted two key mechanisms underlying these improvements: the enhanced recognition of long‐term hydrological patterns through improved memory and a better representation of catchment heterogeneity by emphasizing non‐climatic attributes. These findings demonstrate that integrating HS offers a superior approach to traditional point‐error‐based calibration in AI‐driven hydrological modeling.

The efficiency and accuracy of grid-based computational fluid dynamics methods are strongly dependent on the chosen cell size. The computational time increases exponentially with decreasing cell size. Therefore, a grid coarsing technology without apparent precision loss is essential for various numerical modeling methods. In this article, a physical adaption neural network (PANN) is proposed to optimize coarse grid representation from a fine grid. A new convolutional neural network is constructed to achieve a significant reduction in computational cost while maintaining a relatively accurate solution. An application to numerical modeling of dynamic processes in landslides is firstly carried out, and better results are obtained compared to the baseline method. More applications in various flood scenarios in mountainous areas are then analyzed. It is demonstrated that the proposed PANN downscaling method outperforms other currently widely used downscaling methods. The code is publicly available and can be applied broadly. Computing by PANN is hundreds of times more efficient, meaning that it is significant for the numerical modeling of various complicated Earth-surface flows and their applications.

Runoff forecasting is a long-standing challenge in hydrology, particularly in unmeasured catchments and rapid flash flood prediction. For unmeasured catchment forecasting, we introduce the encoder-decoder-based dual-layer long short-term memory (ED-DLSTM) model[1]. This model fuses static spatial granularity attributes with temporal dynamic variables to achieve streamflow forecasting at a global scale. ED-DLSTM reaches an average Nash efficiency coefficient (NSE) of 0.75 across more than 2000 catchments from historical datasets in the United States, Canada, Central Europe, and the United Kingdom. Additionally, ED-DLSTM is applied to 150 fully ungauged catchments in Chile, achieving a high NSE of 0.65. The interpretability of the transfer capacities of ED-DLSTM is effectively tracked through the cell state induced by adding a spatial attribute encoding module, which can spontaneously form hydrological regionalization effects after performing spatial coding for different catchments.Moreover, rapid flood prediction with daily resolution is challenged to capture changes in runoff over short periods. To address this, we also propose a benchmark evaluation for runoff and flood forecasting based on deep learning (RF-Bench) at an hourly scale. We introduce the Mamba model to hydrology for the first time. The benchmark also includes Dlinear, LSTM, Transformer, and its improved versions (Informer, Autoformer, Patch Transformer). Results indicate that the Patch Transformer exhibits optimal predictive capability across multiple lead times, while the traditional LSTM model demonstrates stable performance, and the Mamba model strikes a good balance between performance and stability. We reveal the attention patterns of Transformer models in hydrological modeling, finding that attention is time-sensitive and that the attention scores for dynamic variables are higher than those for static attributes.Our work [2,3] provides the hydrological community with an open-source, scalable platform, contributing to the advancement of deep learning in the field of hydrology. [1] Zhang, B., Ouyang, C., Cui, P., Xu, Q., Wang, D., Zhang, F., Li, Z., Fan, L., Lovati, M., Liu, Y., Zhang, Q., 2024. Deep learning for cross-region streamflow and flood forecasting at a global scale. The Innovation 5, 100617. https://doi.org/10.1016/j.xinn.2024.100617[2] Zhang, B., Ouyang, C., Wang, D., Wang, F., Xu, Q., 2023. A PANN-Based Grid Downscaling Technology and Its Application in Landslide and Flood Modeling. Remote Sensing 15, 5075. https://doi.org/10.3390/rs15205075[3] Xu, Q., Shi, Y., Bamber, J.L., Ouyang, C., Zhu, X.X., 2024. Large-scale flood modeling and forecasting with FloodCast. Water Research 264, 122162. https://doi.org/10.1016/j.watres.2024.122162

Timely recognition of debris-flow-generated seismic signals is essential for event-based warning, yet many machine-learning workflows rely on handcrafted features that may discard task-relevant information before model training. Here we develop and evaluate a raw-waveform-based representation-learning framework for debris-flow waveform classification and downstream warning. The proposed Patch Fourier Transformer (PFT) converts three-component seismic waveforms into localized short-time-Fourier-transform patches and uses attention-based sequence modeling to learn the temporal evolution of flow-generated seismic energy. The model was trained, tested, and evaluated using seismic records from the Illgraben catchment in Switzerland, with the 2020 events reserved for retrospective continuous-warning simulation. The PFT achieved F1 scores above 0.96 at all three Illgraben stations and issued network-level warnings 24–170 min before debris-flow fronts reached the downstream check dam CD29. No event-level missed detections or false alarms were observed in the 2020 retrospective simulation. Compared with feature-based machine-learning models and an STA/LTA baseline, the PFT provided more stable waveform classification and warning performance under continuous field conditions. Preliminary tests on five unseen catchments further suggest potential transferability of the learned waveform representations, although site-specific validation remains necessary. These results indicate that raw-waveform-based time–frequency representation learning can reduce dependence on handcrafted seismic features and support more reliable event-based debris-flow warning.

Seismic instruments can record a wide range of Earth’s surface mass movements processes, and machine learning techniques have been applied to the detection of seismic signals generated by debris flows for early warning purpose. However, several challenges remain, including but not limited to the following: 1) Traditional models such as Random Forests are unable to directly process raw waveform signals and rely heavily on manual feature extraction, which can introduce redundant or insufficient features, leading to unreasonable inductive bias; 2) Conventional time-series models suffer from limited interpretability when input with raw waveform. To address these limitations, this study presents an improved model based on the Transformer architecture, referred to as the Patch Fourier Transformer (PFT). The PFT model overcomes the limitations of manual feature extraction by directly processing raw waveform data. Furthermore, the incorporated patch attention mechanism effectively captures energy variations in the time-frequency domain and produces attention weights that closely align with the spatio-temporal distribution of the spectrogram. Using seismic data from the 2020 Illgraben debris flow event for testing, the PFT model achieved over 96% accuracy in waveform identification. Additionally, the early warning system provided downstream (CD29) alerts ranging from tens of minutes to several hours without any false alarms. Our proposed framework simplifies the debris flow detection process, provides a degree of interpretability while validating cross-region detection performance, and lays the foundation for dealing with more complex scenarios in the future.