This study proposes the “Ina-EcoShield-CoastalGuard” framework, a data-driven coastal erosion susceptibility framework that uses machine learning to estimate the probability of coastal erosion and to test whether coastal vegetation contributes measurably to that probability along the wave-dominated Southern Coast of the Special Region of Yogyakarta (DIY), Indonesia. The framework integrates 14 indicators from seven variable groups (hydrodynamics, vegetation, morphology, morphodynamics, geo-environment, climate, and human activities) derived from open geospatial products, complemented by three detailed vegetation metrics from Sentinel-2 and ESA WorldCover. Variance Inflation Factor (VIF) analysis was used to screen multicollinearity, and the effect of variable selection was tested against the full and PCA-based variable sets. Three tree-based algorithms (Random Forest, RF; Gradient Tree Boosting, GTB; Classification and Regression Trees, CART) and two ensembles (soft voting and stacking) were trained on 60 field-verified reference points (32 eroding, 28 stable or accreting) and evaluated with nested, repeated 5-fold cross-validation, spatial block cross-validation, and an independent validation against Digital Shoreline Analysis System (DSAS) end-point rates (EPR). The ensembles achieved the highest discrimination (stacking AUC 0.971 ± 0.035; voting 0.970 ± 0.034; RF 0.959 ± 0.047). Without the shoreline-change predictor, models trained on the field points still separated eroding from non-eroding shoreline pixels in the independent EPR data (RF AUC 0.95, 95
To address the problems of discrete TOC calibration samples, continuous well-log responses, strong vertical heterogeneity, and complex responses in organic-rich intervals in the ER well-series logging data from the KAS Block of Daqing Oilfield, this study proposes a well-log TOC reconstruction and grade identification method based on geological frequency-aware attention and retention network, named GFAR-TOCNet. The method takes multi-curve logging data, including GR, RT, AC, DEN, CNL, SP, and CAL, as input, and constructs samples using a sliding-window strategy with center-point supervision, thereby establishing an effective correspondence between continuous logging responses and discrete TOC calibration values. In the model architecture, multi-scale dilated convolution is first used to extract thin-bed boundaries, local textures, and logging responses at different scales. A retention network is then introduced to capture long-range depth dependencies and sedimentary continuity. Furthermore, frequency-aware geological attention integrates low-frequency sedimentary background, high-frequency boundary disturbances, and geological prior responses, improving the model’s ability to identify organic-rich intervals and complex transitional zones. Experimental results show that GFAR-TOCNet outperforms traditional methods and several deep learning models in both continuous TOC reconstruction and six-level TOC grade identification. On the test set, the model achieves an R2 of 0.934 and a six-level identification accuracy of 91.42
Rapid drawdown is a transient hydro-mechanical loading condition in which external water support is removed faster than pore-water pressures can dissipate within low-permeability zones of an earth dam. This study presents a coupled numerical assessment of a 71.95 m high earth dam in central Nepal using transient seepage analysis, B-bar-stabilized finite element strength reduction, and multiple limit-equilibrium formulations. The dam has a 285 m base width and a normal upstream water depth of approximately 49 m. A 40-day transient drawdown was considered, while instantaneous undrained unloading was treated separately as a conservative limiting condition rather than as a substitute for the transient analysis. Under the full-reservoir steady-state condition, the nine limit-equilibrium formulations considered in the original model produced factors of safety between 1.856 and 2.234, with a mean of approximately 2.00. Under the instantaneous post-drawdown limiting condition, the LEM factors of safety ranged from 1.701 to 1.934, whereas the FEM strength-reduction analysis produced a critical SRF of 1.35 and a maximum displacement of 0.455 m. The lower FEM SRF is interpreted as evidence of progressive deformation and strain localization that are not represented explicitly in rigid-body LEM formulations. A model-form reliability analysis based on the dispersion among the nine LEM results produced β = 8.26 before drawdown and β = 10.12 for the instantaneous post-drawdown LEM set. These values do not represent system failure probabilities because material, permeability, groundwater, spatial, and parameter uncertainties were not sampled; they quantify only inter-method dispersion and are therefore reported as model-form indicators. The study demonstrates that transient pore-pressure redistribution, loss of external hydrostatic support, and progressive strain localization must be considered jointly when evaluating rapid drawdown. The framework is directly applicable to earth dams and, with recalibration of constitutive and hydraulic properties, can also inform analogous assessments of water-retaining mining embankments.
Accurate predictive modeling of groundwater hydrochemical parameters is vital for the sustainable management of drinking and irrigation water resources. This study develops a robust machine learning framework by coupling Extreme Gradient Boosting (XGB) with Enzyme Activation Optimization (EAO) to predict electrical conductivity (EC) and total hardness (TH) in the unconfined Talesh aquifer, northern Iran. Preprocessing incorporated quality screening, systematic data imputation, and Pearson correlation-based feature selection combined with domain-specific hydrochemical considerations to identify the most influential predictor ions. The proposed EAO-XGB model significantly outperformed conventional baseline models (including standalone XGB, Support Vector Machines, Artificial Neural Networks, and Linear Regression), achieving superior predictive accuracy on the independent test dataset for both EC (R2 = 0.910, RMSE = 64.76 μS/cm, MAE = 27.73 μS/cm) and TH (R2 = 0.919, RMSE = 25.84 mg/l, MAE = 8.70 mg/l). Model reliability and generalization stability were rigorously verified through a fivefold cross-validation uncertainty analysis. SHapley Additive exPlanations (SHAP) applied to the optimized EAO-XGB model revealed that major divalent cations (Ca2+ and Mg2+) dominantly control TH dynamics in accordance with stoichiometric principles, while salinity variations (EC tolerations) are primarily driven by dominant mobile electrolytes. Hydrochemical trend assessments indicate that predicted EC values predominantly range within 0–1000 μS/cm (reflecting fresh to moderately mineralized groundwater) and TH levels remain between 0–500 mg/l (as CaCO3), complying with World Health Organization (WHO) permissible guidelines for potable and domestic usages without imposing severe salinization hazards. The proposed framework demonstrates high fidelity for high-resolution groundwater quality assessment in coastal and alluvial aquifers.
Rock permeability plays a crucial role in characterizing subsurface fluid flow, while pore-scale numerical simulations such as the Lattice Boltzmann Method (LBM) can be computationally expensive when applied to large numbers of digital rock samples. To address this challenge, this paper introduces a channel-attention-based convolutional neural network that synergistically integrates Squeeze aggregated Excitation (SaE) modules with dense connectivity. This architecture enhances feature representation by autonomously emphasizing informative channels and effectively fusing multi-scale spatial information, thereby improving the representation of complex pore structures. The model was trained on a digital rock dataset where permeability values were obtained from LBM simulations as the numerical reference, and its performance was rigorously evaluated through comprehensive experiments. Results demonstrate that the model achieves competitive prediction accuracy, outperforming other popular 3D networks, while maintaining favorable computational efficiency. Additionally, the model demonstrates favorable transfer-learning performance on additional sandstone datasets with limited target-domain data. This combination of predictive performance, efficiency, and data adaptability underscores the model’s practical value in scenarios where data availability is limited, offering an efficient surrogate for LBM-based permeability prediction in digital rock analysis.
