Accurate real-time estimation of the tropical cyclone (TC) intensity is crucial for TC forecasting and disaster management. Conventional intensity estimation methods like the Dvorak technique are subjective, and most deep learning (DL) approaches ignore the natural variability and time lag between the actual intensity and the remote sensing cloud pattern imagery. This paper introduces a multi-view sequence fusion network (MSFN), which is a novel DL framework that enhances TC intensity estimation by combining multi-channel cloud pattern imagery, sea surface temperature (SST), and statistical factors. The MSFN features three key innovations: a rotation-equivariant convolutional neural network (CNN) encoder to extract robust cloud pattern features, a transformer-based feature interactor to capture both temporal and cross-view dependencies, and a Gated Memory Fusion (GMF) module to dynamically integrate multi-view data and enhance robustness against noise. Tested on the full 2016–2019 TC data over the western North Pacific (WNP) basin, the MSFN achieves a coefficient of determination (R²) of 0.845 and a root mean square error (RMSE) of 4.6–5.2 m s⁻¹. When further evaluated on a strictly homogeneous subset of the same dataset, the MSFN achieves an R² of 0.848 and an RMSE of 4.3 m s⁻¹. Statistical analyses demonstrate that MSFN significantly outperforms the Advanced Dvorak Technique (ADT) and achieves comparable overall performance to the Satellite Consensus (SATCON) algorithm, with superior accuracy for weaker systems below 45 knots ( = 23.1 m s⁻¹). Interpretability analyses and ablation studies confirm the model’s focus on meteorologically significant features and its own component synergy. The MSFN offers a reliable, interpretable tool for objective TC intensity estimation, advancing the TC forecasting capabilities.

Numerical models face persistent challenges in subseasonal-to-seasonal (S2S) precipitation forecasting over China due to strong precipitation variability and complex multi-sphere coupling at S2S timescales. In recent years, artificial intelligence (AI)-based post-processing has emerged as a promising approach, owing to its capacity to learn complex nonlinear relationships and correct systematic model biases from historical data. However, most existing AI-based methods neglect the spatial structure and physical interactions among multi-sphere predictors (e.g., atmosphere, ocean, and land), limiting their ability to capture the underlying dynamics required for physical consistency. This study develops an S2S precipitation bias-correction network (S2SPre-BCNet) based on a cycle-consistent generative adversarial network (CycleGAN), which incorporates causality-selected multi-sphere predictors as conditional inputs to improve weekly accumulated precipitation forecasts from the ECMWF S2S system over China at lead times of 1–6 weeks. Compared to the ECMWF S2S, S2SPre-BCNet reduces mean RMSE (root mean square error) by 11.6% (maximum 17.2%), increases mean ACC (anomaly correlation coefficient) by 27.2% (maximum 49.2%), and raises mean HSS (Heidke skill score) by 1.23% (maximum 2.12%). Across the case studies, S2SPre-BCNet lowers the absolute mean precipitation error by 16.4%. Additionally, interpretability analyses reveal that multi-sphere predictors contribute distinctly across lead times, and the model focuses on physically meaningful regions where precipitation dynamics are most complex, highlighting the potential of causality-informed AI for operational S2S bias correction. This study underscores that AI techniques augmented by causality-based predictor selection can effectively correct biases in forecasts produced by numerical models, enabling their use in operational forecasting.

Tropical cyclones (TCs) are one of the most frequent disastrous weather events in China, causing widespread damage. Traditional approaches for assessing TC disaster damage treat the TC affected regions as isolated units and ignore inter-regional interactions, resulting in underestimation of complex dynamics in disaster damage assessment. In this paper, we developed an original TC disaster damage dataset, with each sample representing a unique disaster event, incorporating city-specific multi-dimensional features and damage indicators. Then, using provincial administrative divisions in China as examples, we innovatively assigned cities as nodes and constructed inter-city interaction graphs. To align with the physical interactions, a deep learning model named TC-Damage is specifically established. It includes an edge building module and a backbone. The edge building module aims to construct inter-city interaction features from multiple perspectives. The backbone employs a multi-layer Graph Neural Network (GNN) based on Graph Sample and Aggregate (GraphSAGE) and Jumping Knowledge Network (JKNet) to learn comprehensive and hierarchical features of inter-city interactions. A loss function combined with focal loss and node-level loss is proposed to address data imbalance and to enforce representation node distribution. Multiple experiments demonstrate that TC-Damage outperforms other assessment methods and effectively identifies high-contribution factors. Explainability analysis of Super Typhoon Lekima reveals that key edges are adjacent to cities with high disaster factors and social development levels and significantly overlap with edges exhibiting strong inter-city interactions.

The North Atlantic Oscillation (NAO) is a major atmospheric mode in the Northern Hemisphere, characterized by frequent fluctuations in sea level pressure (SLP) across the North Atlantic sector. In the development and evolution of the NAO, various dynamic physical processes such as the El Niño–Southern Oscillation (ENSO) and Madden–Julian Oscillation (MJO) influence it to different extents. Previous studies using numerical models or deep learning models for daily NAO forecasts have not accounted for the impact of these dynamic physical processes, making accurate and stable NAO forecasting still a challenge. In this study, the Varimax-Rotation Principal Component Analysis (PCA) and data-driven causal inference are used to identify key dynamic physical processes linked to the NAO. Based on these, a deep learning model called the NAO-Causal Weighted Model (NAO-CWM) is developed, which incorporates causal relationships to assign different weights to these processes, providing effective daily forecasts with a lead time of 1–14 days. Evaluation results show that NAO-CWM outperforms the advanced numerical models, offering reliable NAO forecasts and a better capturing of NAO variation trends.