Reconstructing high‐resolution sea surface temperatures (SST) from staggered SST measurements is essential for analyzing earth system processes. However, when SST measurements are sparse, the resulting inferred SST fields are rather inaccurate. Here, we show that Sparse Discrete Empirical Interpolation Method (S‐DEIM) can be used as a model‐free data assimilation method to reconstruct high‐resolution SST fields from sparse in situ observations. The S‐DEIM estimate consists of two terms, one computed from instantaneous in situ observations using empirical interpolation, and the other learned from the historical time series of observations using recurrent neural networks (RNNs). We train the RNNs using the National Oceanic and Atmospheric Administration's weekly high‐resolution SST data set spanning the years 1989–2021 which constitutes the training data. Subsequently, we examine the performance of S‐DEIM on the test data, comprising January 2022 to January 2023. For this test data, S‐DEIM infers the high‐resolution SST from 100 in situ observations, constituting only 0.2% of the high‐resolution spatial grid. We show that the resulting S‐DEIM reconstructions are about 40% more accurate than earlier empirical interpolation methods, such as DEIM and Q‐DEIM. Furthermore, 91% of S‐DEIM estimates fall within C of the true SST. We also demonstrate that S‐DEIM is robust with respect to sensor placement: even when the sensors are distributed randomly, S‐DEIM reconstruction error deteriorates only by 1%–2%. S‐DEIM is also computationally efficient: training the RNN, which is performed only once offline, takes approximately 1 minute. Once trained, the S‐DEIM reconstructions are computed in less than a second.
Abstract Physics‐based handcrafted features, such as P‐wave amplitude derived from early parts of active source ultrasonic waveforms, successfully predict shear stress evolution. This study investigates whether feature extraction can be automated by leveraging full active source ultrasonic waveforms recorded during a laboratory friction experiment. We compare four models: handcrafted feature‐driven model (reference), convolutional neural network (CNN) model, explainable data‐driven (XDD) model, and a model with an explainable generalized physics‐based (XPG) feature. The reference model uses the spectral amplitude of the P‐wave packet calculated at center frequency, while the CNN relies on convolutional and pooling layers for feature extraction. The XDD model combines CNN feature extraction with long short‐term memory units and an attention layer to capture feature time history and significance. The novel XPG extracts a generalized spectral amplitude feature through a custom multilayer perceptron designed to learn linear combination of signals' spectral amplitudes without prescribing specific frequencies. Although CNN‐extracted features resemble handcrafted ones, both the CNN and XDD models underperform compared to the reference, while the XPG model outperforms all others. The attention mechanism in the XDD model highlights the importance of first P‐wave packet, whereas the XPG model also identifies later‐arriving shear wave packet as significant. The prediction improvement by inclusion of shear waves is physically intuitive; S‐waves have particle motion polarized in the direction of fault shear and are likely more sensitive to frictional state than P‐waves. This work demonstrates the value of incorporating domain knowledge into automatic feature extraction with potential applications in seismic data analysis.
Abstract Accurate spatiotemporal wind speed prediction is essential for improving wind energy utilization and ensuring power grid security. However, existing methods often rely on incorporating multiple meteorological variables, which increases model complexity and obscures the role of large‐scale circulation in guiding local wind evolution. This study proposes a physics‐aware gated spatiotemporal fusion network (PAG‐STFN) for 24‐hr wind speed prediction over southern China and its coastal regions. By introducing mean sea level pressure to represent the large‐scale circulation background, the model adaptively modulates historical wind field features. Results show that PAG‐STFN yields smaller mean errors than the baseline models, with particularly clear advantages at lead times of 6 hr and beyond. The model also preserves the seasonal wind field structures over the study region and demonstrates good predictive performance at offshore locations near Jinwan and Changle. Analysis of a cold wave event further illustrates that PAG‐STFN can adaptively regulate the intensity and spatial location of the injected background information, thereby contributing to improved prediction accuracy. Meanwhile, PAG‐STFN introduces only a marginal increase in model parameters while maintaining fast inference, highlighting its practical potential in wind power prediction and operational dispatch.
Abstract During rocket launches, atmospheric disturbances significantly impact the ionosphere, generating Traveling Ionospheric Disturbances (TIDs). These disturbances can affect radio communications and navigation systems and may pose risks to high‐speed aerospace vehicles. By analyzing TIDs induced by SpaceX rocket launches, this study aims to understand the ionospheric response to rocket launches and explore potential high‐speed flight trajectories. This work investigates 153 Falcon rocket launch events from January 2023 to June 2024. First, ionospheric data undergoes preprocessing, where the Savitzky‐Golay filter and Butterworth filter are applied to extract ionospheric disturbances caused by rocket launches, and disturbance images are subsequently generated. Next, the occurrence time, duration, and spatial extent of these disturbances are analyzed, and TID images are manually labeled. Finally, using the labeled TID images, a deep learning‐based ionospheric disturbance detection model is trained with a target detection algorithm. This model achieves an identification accuracy of approximately 83% in detecting TIDs within TEC images.
Abstract Rapid and reliable coastal inundation mapping from Synthetic Aperture Radar (SAR) imagery is vital for disaster response. Although recent deep learning advances have enhanced this task's accuracy and automation, two major challenges remain: (a) geographic and imaging variations across flood events limit the generalizability of supervised models trained on static data sets, and (b) constructing large‐scale annotated flood data sets remains costly and time‐consuming. These challenges are especially pronounced in coastal regions, where precise segmentation is critical but conventional models often struggle to adapt to unseen scenarios, leading to degraded performance. To address this, we propose Coastal Inundation Mapping via Test‐Time Domain Adaptation (CIM‐TTDA) for SAR‐based coastal flood mapping. Instead of requiring new labels, CIM‐TTDA adapts a pretrained water segmentation model to each unseen event using only unlabeled post‐event SAR imagery. The core of CIM‐TTDA is a self‐supervised learning strategy tailored for coastal inundation mapping. Built upon pseudo‐label learning and exponential moving average updates, the strategy ensures stable adaptation across diverse flood events while maintaining computational efficiency. We evaluate CIM‐TTDA on three cyclone‐induced flood events (eight distinct scenes) and conduct five comprehensive experiments to assess its effectiveness, robustness, and efficiency. Results show that CIM‐TTDA improves pixel‐level mean Intersection over Union (IoU) by approximately 0.15 over conventional supervised models across various convolutional‐ and Transformer‐based architectures. With the EfficientNet‐B0 architecture as the pretrained model, CIM‐TTDA achieves a mean IoU of 0.803, outperforming supervised baselines and state‐of‐the‐art test‐time domain adaptation methods. Overall, CIM‐TTDA offers a promising solution for accurate and stable coastal inundation mapping in unseen real‐world scenarios.
Abstract Effective properties describe transport processes in porous media, controlled by their geometric characteristics (i.e., pore‐size distribution, connectivity, and porosity) and phase configuration. Estimating effective properties of partially saturated porous media requires performing computationally expensive pore‐scale simulations. This process typically involves two sequential simulations that are high dimensional and require long simulation times. To address this challenge, machine learning methods are combined with physics‐based simulations to enable fast initialization. We propose the saturation‐conditioned U‐Net method for efficiently predicting initial phase configuration in pore‐network structures. This approach employs a shallow neural network to condition the U‐Net model via skip connections. We demonstrate the performance of this method using two pore‐network structures under different water saturation conditions. The method successfully delivers accurate predictions while maintaining reasonable computational resources requirement. It achieves nearly 8×103 times speed‐up compared to Lattice‐Boltzmann simulations for predicting gas phase distribution in a 3D structure. When coupled with a physics‐based solver, this approach enables efficient and accurate calculation of effective properties of partially saturated media.
Abstract Chlorophyll‐a (Chl_a) concentration in the ocean is a critical indicator of primary production and plays a pivotal role in the global carbon cycle. Its accurate predictions are essential for reliable projections of future climate change. However, Chl_a is controlled by myriads of interactive physical and biogeochemical processes, posing a substantial challenge for its diagnosis and prognosis through numerical modeling. Machine learning (ML) approaches offer a promising venue to face the challenge, yet existing ML models often struggle in handling Chl_a's significant spatiotemporal variability. We propose a novel machine learning model, Spatiotemporal Dynamics Hunter (STD‐Hunter), to tame spatiotemporally heterogeneous dynamics of Chl_a and improve its prediction. STD‐Hunter effectively accommodates the heterogeneity in principal dynamics characterized by multiple basis ML models. By integrating these basis models with Chl_a's spatiotemporal characteristics, STD‐Hunter forms task‐specific predictive models that harvest the underlying dynamics. STD‐Hunter optimizes the trade‐off between effectively leveraging data and preserving idiosyncratic dynamics. Based on remotely sensing data, we demonstrate the superior performance of STD‐Hunter in predicting highly variable Chl_a in an active marginal sea. The extracted characteristics are also well aligned with Chl_a's intrinsic dynamics. By bridging between highly nonuniform observations and their underlying dynamics, STD‐Hunter offers an interpretable tool to obtain physically meaningful spatiotemporal characteristics from a data‐driven perspective.
Abstract Electrical resistivity tomography (ERT) is a subsurface imaging geophysical technique. Traditional ERT inversion methods, such as smoothness‐constrained least‐squares approaches, often suffer from discretization artifacts when reconstructing electrical resistivity from resistance measurements. To address the limitations of conventional ERT inversion, this study introduces a novel super‐resolution framework based on an Implicit Neural Representation (INR) to enhance 3D ERT inversion resolution beyond that achievable with standard Gauss‐Newton techniques. The proposed machine learning methodology effectively integrates high‐resolution two‐dimensional data with coarse three‐dimensional data to generate a higher‐resolution resistivity representation consistent with the measured apparent‐resistivity data. Validation using data from the Waste Isolation Pilot Plant (WIPP) site shows that the INR approach improves 3D inversion quality relative to Res3Dinv. For the WIPP data set, the INR model increases R 2 from 0.107 to 0.361, reduces RMSE from 77.21 to 33.32 Ω·m, and reduces bias from 68.34 to 13.94 Ω·m (corresponding to relative improvements of 237.4%, 56.8%, and 79.6%, respectively). Tests on a synthetic domain with increased resistivity variation show even stronger performance enhancement: R 2 increases from 0.149 to 0.740, RMSE decreases from 119.28 to 53.24 Ω·m, and bias decreases from 88.45 to 12.74 Ω·m (relative improvements of 396.6%, 55.4%, and 85.6%, respectively). Results demonstrate improved apparent‐resistivity prediction using INR, but the method is deterministic and does not provide formal uncertainty bounds. Generalization to other electrode arrays and systematic quantification of line‐density sensitivity remain future work.
Abstract Achieving both accuracy and interpretability in deep learning models for geochemical anomaly recognition constitutes a significant challenge. To overcome this challenge, this study developed a novel interpretable dual‐branch network combining a spectral attention bidirectional RNN (BiRNN) branch and a spatial attention CNN branch guided with geological knowledge for geochemical anomaly recognition. The dual‐branch network comprehensively extracts rich spectral and spatial features from geochemical data cubes, which record both the elemental composition within individual pixels and the spatial relationships between neighboring pixels. To mitigate deep learning's black‐box nature and enhance interpretability, geological knowledge (e.g., key ore‐controlling faults) and attention mechanisms are incorporated into the model, boosting both predictive accuracy and interpretability. Geological knowledge integrated during training constrains the model, penalizing predictions that violate geological principles. Concurrently, the attention mechanisms applied to both branches enable visual interpretation of the decision‐making process. The model's efficacy for identifying mineralization‐associated geochemical anomalies was validated in a case study within the southern Tianshan, China. Ablation experiments focusing on pre‐model interpretability via geologically constrained input, in‐model interpretability via incorporating geological knowledge into the loss function of the model, and post‐model interpretability via visualization of attention weights are conducted. Comparative results show that the geologically constrained dual‐branch network improves both accuracy and interpretability in geochemical anomaly detection. Furthermore, spectral and spatial weight attention visualizations further demonstrate the model's alignment with geological principles.
Abstract Full waveform inversion (FWI) reconstructs subsurface models by minimizing the mismatch between observed and simulated seismic data. However, the strong nonlinearity of the inversion problem makes gradient‐based optimization highly sensitive to inaccurate initial models and incomplete observational data. These challenges become even more pronounced for vertical seismic profiling (VSP) data due to limited acquisition geometry and sparse receiver distribution. Recent studies suggest that neural network‐based reparameterized FWI can improve inversion robustness and reduce the dependence on high‐quality initial models. Nevertheless, the mechanisms underlying this improvement remain insufficiently understood. In this study, we investigate the robustness of reparameterized FWI from the perspective of loss landscape geometry through loss visualization. The analysis reveals how neural network parameterization reshapes the optimization landscape, enabling stable inversion under challenging conditions, including inaccurate initial models, missing low‐frequency information, and low signal‐to‐noise ratio data. Based on these insights, we propose a physics‐informed Transformer‐based reparameterized FWI framework (PIT‐FWI) for VSP velocity inversion. Synthetic experiments using the Sigsbee model demonstrate that the proposed method significantly improves inversion accuracy and stability compared with conventional FWI and UNet‐based reparameterized FWI under degraded observational conditions, highlighting its robustness in challenging inversion scenarios and its potential for future application to field VSP data.
Abstract Precipitation nowcasting refers to the high‐resolution forecasting of rainfall and hydrometeors within 0–6 hr according to the official definition of the World Meteorological Organization, which has relied on numerical models for decades. Recently, artificial intelligence (AI) has shown promise in addressing precipitation nowcasting. However, three key scientific issues have not been studied: (a) the researcher does not know what semantic knowledge the AI model has learned in the high‐dimensional space that influences precipitation forecasts. (b) If the learned semantic knowledge can be quantified, which aspects of precipitation does each factor control? And (c) can the semantic knowledge contribute to improving the accuracy of precipitation nowcasting? Hence, this study proposes a human supervised spatial‐temporal disentanglement model (STNet) that perturbs high‐dimensional vectors in the latent space and enforces orthogonality constraints to disentangle the spatial‐temporal semantic knowledge learned by the model—referred to as disentangled latent factors (a set of orthogonal high‐dimensional vectors). Qualitative and quantitative experiments on the specialized nowcasting dataset (SEVIR) reveal that the disentangled latent factors primarily fall into two categories: temporal factors, which predominantly control the evolving precipitation pattern (dynamics), and spatial factors, which mainly regulate precipitation intensity (static). Specifically, the temporal factor constraint provides more significant improvements in precipitation area forecasting. The spatial factor constraint, on the other hand, focuses on optimizing precipitation intensity.
Abstract Access to accurate and timely soil moisture information is critical for effective water resource management, ensuring food security, and advancing climate modeling; however, such information is often limited. We propose a novel transferable Fourier Neural Operator (FNO) framework designed to estimate subsurface soil moisture utilizing surface‐level data (5 cm). This framework demonstrates robust generalization capabilities for soil moisture estimates for deeper soil layers, wherein prediction tasks are inherently more complex. The model has been trained on high‐resolution temporal data from six distinct sites encompassing both arid and humid subtropical regions, as well as various soil types. It reliably predicts soil moisture at depths of 15 and 50 cm. By leveraging transfer learning, pre‐training, and fine‐tuning with only 5% of the target domain data, we achieved improved accuracy and stability. This innovative framework provides a scalable and data‐efficient solution for delivering subsurface soil moisture data to users with constrained data sets.
Abstract We demonstrate a data‐driven parameterization framework for snow albedo evolution using a constrained auto‐regressive neural differential equation that directly predicts snow albedo change from standard meteorological inputs. After training with multi‐year in situ and satellite observations from a wide variety of locations, the scheme effectively reproduces daily albedo evolution across diverse climate zones, with median error under 7.5% (RMSE ≈ 0.05), a 10%–30% improvement over established models. Furthermore, the model generalizes to sites not seen during training and scales from coarser grids to point locations. The scheme can easily incorporate new features as observational networks expand, offering an adaptive and computationally lightweight framework for next‐generation land and climate models.
Abstract Marine heatwaves (MHWs) pose increasing risks to marine ecosystems and climate‐sensitive activities. Subseasonal‐to‐seasonal (S2S) MHW forecasts are essential for early warning, yet dynamical prediction systems show limited reliability. Here, we present MHWCorrNet, an interpretable deep learning‐based post‐processing framework for ECMWF IFS S2S MHW forecasts. It improves globally averaged MHW forecast skill by ∼9% at 1–6‐week lead times, with larger gains of ∼15% and ∼22% in the Northern and Southern Hemisphere extratropics. We show that regional background states systematically shape forecast correction strategies, revealing a previously overlooked aspect. In the extratropics, more variable air‐sea processes cause more missed events, so corrections primarily improve event detection. In the tropics, weaker thermal contrasts between MHW and non‐MHW conditions produce more marginal events near the detection threshold; therefore, skill gains mainly result from reduced false alarms. By leveraging interpretable diagnostics to quantify feature importance, MHWCorrNet adaptively modulates its reliance on key physical variables across latitude bands, placing stronger dependence on multiple predictors in extratropical regions to correct large residual errors, while exhibiting weaker dependence in the tropics, where sea surface temperatures variability is constrained by large‐scale climate modes and the correctable error space is smaller. This physically grounded adaptation enhances both forecast reliability and interpretability, underscoring the value of incorporating physical understanding into data‐driven prediction frameworks.
Abstract Ocean models can represent surface circulation at kilometer scales, but their computational cost limits broad experimentation. We present DeepCUN, a deep convolutional encoder–decoder (U‐Net) that emulates daily mean Baltic Sea surface current components on a 1‐nautical‐mile grid. DeepCUN is trained on 2015–2023 reanalysis currents with atmospheric reanalysis forcing. An occlusion‐sensitivity channel ablation indicates that wind provides the dominant predictive information when combined with the antecedent surface‐current state, while additional atmospheric variables contribute negligibly. Accordingly, DeepCUN is configured to use only the prior surface‐current state and the subsequent‐day wind components to predict next‐day surface current fields. On an independent 2024 test year, DeepCUN captures the dominant spatial patterns and temporal variability, with the highest errors concentrated in the southwestern boundary exchange corridor and narrow straits and lower errors across the basin interior, where correlations exceed 0.9 in most locations. For post hoc explainability of this data‐driven mapping, we apply layer‐wise relevance propagation to characterize attributed support and a diagonal Jacobian elasticity metric to quantify local responsiveness, revealing spatially varying reliance on state memory in the southwestern boundary‐influenced region and wind‐modulated adjustment across the basin interior. In addition, to assess the potential forecast skill of this approach, we perform recursive rollouts, which show domain‐mean absolute error increasing smoothly from ∼2.5cms−1 at 1 day to ∼6.0cms−1 at 21 days while correlation decreases from >0.9 to ∼0.65, indicating stable multi‐week behavior under idealized forcing conditions. The framework provides an efficient and interpretable template for configuring, pruning, and stress‐testing reduced‐input emulators of regional surface circulation.
Abstract Mantle convection drives the solid Earth, powering plate motions, volcanism, and earthquakes while regulating planetary heat loss. Reconstructing its history is hampered by sparse, noisy observations concentrated near the surface and the present day. Here I develop an inverse physics‐informed neural network framework to estimate mantle thermal convection as continuous space‐time fields. Twin experiments assimilate irregular near‐surface kinematics from tracer trajectories together with a terminal interior temperature structure, used as a proxy for present‐day imaging. The method reconstructs transient temperature and flow with reasonable accuracy under the tested noise levels. Using either constraint alone yields nonunique or physically distorted histories, demonstrating that complementary surface and terminal information is essential for accurate reconstruction. Training shows staged transitions that illuminate multi‐constraint learning. Overall, this work provides a practical route to integrate heterogeneous geophysical constraints into retrospective reconstructions of mantle dynamics.
Abstract Influenced by tectonic, geophysical, and environmental aspects, the release of carbon dioxide (CO 2 ) in fault systems is a fundamental component of Earth's carbon cycle. Appreciating their contribution to natural greenhouse gas flow depends on knowing these emissions. We integrated 867 degassing records with harmonized raster variables to provide a predictive framework utilizing machine learning to estimate CO 2 fluxes in tectonically active areas. Grid search cross‐validation trained and optimized five regression models: Support Vector Regression, K‐Nearest Neighbors, Decision Tree, Random Forest, and Gradient Boosting. Then, a stacking regressor was used, which achieved an R 2 of 0.677, surpassing the individual models. To avoid extrapolation into tectonically stable regions, the model was applied pixel by pixel inside an active‐tectonic mask that defines the applicability domain of the model, using only globally available raster variables such as Bouguer anomaly, heat flow, peak ground acceleration, mean annual temperature, vegetation cover, soil properties, porosity, and permeability. The resulting map highlights coherent high‐flux zones along the Pacific Ring of Fire, the Andes, and other active orogens, confirming that geodynamic and geothermal processes strongly control CO 2 degassing. The given method provides a repeatable approach for subsequent research on subsurface degassing and its role in the global carbon cycle, thereby improving the capacity to define natural CO 2 emissions in tectonically active areas.
Abstract We present an automated system for identifying minerals and classifying rock types in Apollo lunar mare basalts using scanning electron microscopy (SEM) imagery. Mineral segmentation is based on a U‐Net architecture, supplemented by two scale‐aware models designed to incorporate pixel size information. We find that a single U‐Net without explicit scale input achieves performance comparable to the scale‐aware models. From a total of 17,248 augmented images (i.e., 1,078 unique base images), 12,800 were used for training, and the final model achieved average pixel‐wise accuracies of 0.85, 0.81, and 0.82 on the training, validation, and test sets, respectively. For rock classification, we constructed a separate data set by compiling reported modal mineral abundances from the literature to train both a rule‐based classifier and a Gaussian Naive Bayes model. These classifiers were then applied to modal abundances derived from the segmented images, achieving accuracies of >∼0.78 in distinguishing between ilmenite, pigeonite, and olivine basalts. Our framework enables rapid, scalable, first‐order mineral identification and rock classification of lunar mare basalt petrography. However, it also reveals several limitations, including challenges in identifying minor phases, the need for phase subclassification, and inaccuracies of human bias in training annotations. These limitations underscore the continued importance of expert interpretation for detailed mineralogical and petrological studies. Nevertheless, our system provides a useful baseline and benchmark with applicability to both existing Apollo collections and lunar meteorites, as well as to future returned‐sample missions.
Abstract Tropical instability waves (TIWs) generate mesoscale sea surface temperature (SST) fluctuations in the eastern equatorial Pacific and influence the evolution of the El Niño‐Southern Oscillation (ENSO). Yet satellite‐based prediction of TIW‐related SST anomalies remains largely SST‐centric and makes limited use of sea surface salinity (SSS) observations. Here we show that satellite SSS provides a practically useful constraint on TIW‐related SST forecasts. We develop a dual‐branch deep learning framework trained on Copernicus Marine Service SST and SSS products, in which a ConvLSTM‐UNet3D trend branch represents slowly varying evolution and a frame‐difference branch highlights transient TIW disturbances. A Mean Deviation Loss strengthens learning in regions of strong anomalies, including the TIW triangular zone and the cold tongue. SST is the sole prognostic output, while SSS is prescribed at each rollout step under three protocols with distinct operational meanings: a deployable, leakage‐free SST‐recursive forecast using training‐derived monthly SSS climatology; a diagnostic upper bound using observed SSS at the corresponding forecast dates; and a constant‐field Zero‐map stress test. Compared with the SST‐only configuration, incorporating time‐varying SSS reduces the one‐step 5‐day SST RMSE from 0.30 to 0.27°C, while the deployable climatological‐SSS setting attains 0.28°C, within 0.01°C of the observed‐SSS upper bound at the single‐step horizon and closely tracking it under multi‐step rollouts. Cross‐spectral and time‐lagged regression analyses further reveal that coherent leading SSSA modes can lead SSTA by 5–12 days on TIW timescales, indicating that seasonal‐background and event‐scale SSS carry physically meaningful, complementary predictive information for TIW‐related SST evolution.
Abstract The dynamic link between rockslide failure and rockfall activity remains elusive, primarily due to the challenge of detecting weak signals amidst high noise. We propose a seismic attribute guided deep learning framework that formulates rockfall detection as a time‐series segmentation task, facilitating the precise extraction of rockfall events from continuous monitoring data. Through self‐supervised pre‐training guided by seismic attributes (e.g., Sum of Squared Differences, SSD), our model captures noise‐robust features that substantially enhance its performance under low signal‐to‐noise ratio (SNR) scenarios. Comprehensive experiments show that, compared with the STA/LTA method commonly used in traditional monitoring and two U‐Net baselines based on waveform and spectral representations, the proposed method achieves improved detection performance and identifies more potential rockfall events than those listed in the existing catalog while maintaining physically consistent energy‐scaling and temporal evolution characteristics. Meanwhile, without re‐training, the model effectively reconstructs rockfall activity sequences at the unseen Illgraben rockslide (Switzerland), demonstrating stronger generalization capability than the purely data‐driven deep learning baselines and uncovering a distinct “growth–acceleration–quiescence–burst” evolutionary pattern prior to the main collapse. This seismic attribute guided deep learning approach provides a reliable technical strategy for short‐term rockslide risk assessment.
