The effect of Global Navigation Satellite System (GNSS)-based tropospheric corrections on Interferometric Synthetic Aperture Radar (InSAR)-derived surface deformation estimates was evaluated for the 24 January 2020 Mw 6.8 Elazığ–Sivrice earthquake. Tropospheric delay is a major source of error in InSAR processing and may affect the interpretation of earthquake-induced deformation fields. To assess this effect, ascending and descending Sentinel-1A images were processed using the Sentinel Application Platform (SNAP) within a standard Differential InSAR (DInSAR) workflow. Phase filtering was applied to reduce noise, and phase unwrapping was performed using the Statistical-Cost, Network-Flow Algorithm for Phase Unwrapping (SNAPHU). Tropospheric delays estimated from GNSS observations were applied to the interferometric displacement products and compared with corrections from the Generic Atmospheric Correction Online Service for InSAR (GACOS). The results show that both GNSS- and GACOS-based corrections influence the spatial distribution of line-of-sight (LOS) displacement fields. However, their effects are not spatially uniform across the study area. Local discrepancies between the corrected products are likely related to regional atmospheric variability, topography, acquisition geometry, and the distribution of GNSS stations. Neither correction strategy produced a spatially uniform improvement across all interferometric pairs; instead, the results highlight the need to carefully evaluate atmospheric correction performance using spatial statistics, profile comparisons, and displacement-difference analyses. This study demonstrates that GNSS- and GACOS-based corrections do not uniformly improve displacement products but provide complementary information for identifying atmospheric contributions and assessing uncertainty in InSAR-derived earthquake deformation fields.
Accurately recognizing mineralization-related geochemical anomalies is essential for mineral exploration targeting, yet it remains challenging in areas characterized by complex geological settings, class imbalance and sparse labeled data. To address these issues, a self-training framework based on Light Gradient Boosting Machine (LightGBM) was proposed for recognizing mineralization-caused geochemical anomalies. In this framework, LightGBM is utilized to capture the nonlinear relationship between the concentration of elements and the spatial distribution pattern of mineral deposits; the self-training algorithm is employed to model the sparse labeled data and vast unlabeled data; and Synthetic Minority Oversampling Technique (SMOTE) is introduced to enlarge the number of minority-class samples so that there are enough positive samples for establishing the LightGBM model during the initial self-training stage. A case study was completed in the Moridawa area of Inner Mongolia, China, to evaluate the applicability of the proposed framework in the recognition of mineralization-caused geochemical anomalies. The self-training LightGBM, self-training support vector classification (SVC), LightGBM, and SVC models were established on the interpolated 1:50,000-scale stream sediment geochemical data, and the performance of the four models was systematically compared in the recognition of mineralization-caused geochemical anomalies. The results show that, among the four models evaluated in the present case study, the self-training LightGBM model achieved the most favorable overall balance among classification performance, spatial prediction efficiency, and computational efficiency. The geochemical anomalies recognized by the self-training LightGBM model cover 2.74
Porosity estimation is one of the most important tasks in sandstone reservoir characterization; however, each method refers to a different pore scale. Laboratory porosimetry provides accurate reference values for core plug pore parameters; however, these discrete values do not adequately capture the spatial variability of pore fabrics. The non-destructive quantification of visible pore space in thin-section images can be performed using Digital Image Analysis (DIA). However, provides a 2-dimensional geometric porosity value (φimage) that cannot easily be used to address issues of pore access, pore continuity, and pore restriction by clay. This study proposes a hybrid Digital Image Analysis–Fuzzy Inference System (DIA–FIS) approach that connects calibrated image measurements and transparent petrophysical reasoning. Pre-fixed hue–saturation–value (HSV) thresholds were used to segment calibrated blue epoxy-impregnated sandstone thin sections, and stability tests were conducted by applying perturbations of ± 5
Lithology identification is a fundamental task in geological exploration and engineering practice, where classification accuracy largely depends on the model’s ability to capture both local texture details and global structural patterns from rock images. Conventional convolutional neural networks (CNNs), although effective in local feature extraction, are limited in modeling long-range spatial dependencies and often exhibit insufficient representation capacity for complex lithological patterns. To address these limitations, this study proposes a novel lithology classification framework, termed SpecMamba-Net, which integrates multi-scale frequency-domain enhancement with selective state-space modeling. Specifically, a spectral enhancement branch is designed to transform feature representations into the frequency domain and emphasize discriminative texture-related spectral components of different lithologies. In parallel, a state-space scanning branch is introduced to model global spatial dependencies and long-range mineral arrangement patterns with linear computational complexity. A gated fusion mechanism is further employed to adaptively integrate spectral features and spatial evolution features, enabling more comprehensive lithology representation. Experiments were conducted on a seven-class lithology dataset using five repeated stratified holdout splits. SpecMamba-Net achieved a mean test accuracy of 0.8654 ± 0.0155 and a macro-F1 score of 0.8667 ± 0.0179, representing the highest mean performance among the evaluated models. Paired split-level comparisons showed statistically supported improvements over most baseline models, although the difference from ResNet18 did not reach the conventional significance threshold. Grad-CAM analysis further provided qualitative evidence that the proposed model attended to lithologically relevant texture and structural regions. These results indicate that SpecMamba-Net provides a promising framework for lithology image classification while maintaining a favorable balance between classification performance and computational efficiency. 1. A novel SpecMamba-Net is proposed in this study to achieve accurate and efficient lithology classification from rock images. 2. A dual-branch architecture is designed consisting of frequency-domain enhancement and selective state space scanning for comprehensive lithological feature extraction. 3. An adaptive gated fusion mechanism is adopted to integrate spectral features and spatial evolution features adaptively. 4. It is verified by comparative experiments and interpretability analysis that the proposed model outperforms mainstream CNNs with higher accuracy and more focused feature representation.
Lokop is a geothermal potential located on the island of Sumatra, Indonesia. It is readily accessible and surrounded by a large population and many emerging business districts. This setting highlights the significant economic potential of the Lokop geothermal field. The government of Indonesia has also pledged a strong commitment to developing green and renewable energy initiatives, including geothermal as one of its priorities. These favorable conditions are ideal for developing a geothermal power plant in the Lokop field. Despite these benefits, the geological framework and geophysical properties of the Lokop prospect remain insufficiently constrained, and published scientific studies on this area remain markedly limited. This research aims to address the existing knowledge gap and was conducted to achieve two main objectives.: (1) to map hydrothermal alteration zones within the Lokop geothermal prospect; and (2) to map fault zones across the study area. In this study, gravity data were used to evaluate fault zones and fracture systems within the Lokop geothermal prospect. Landsat data were used to map the hydrothermal alteration zones. The gravity-based analysis successfully showed the spatial distribution of faults and fracture systems across the Lokop field. Low-gravity Bouguer values (-20 to -15 mGal) were observed within the Lokop field and are interpreted as indicative of porous rocks containing hydrothermal fluids. Furthermore, the Landsat analysis showed the presence of hydrothermal alteration zones, with the altered minerals-primarily iron oxides, kaolinite, and chlorite-identified mainly in the hot spring areas. Time-series evaluation of surface temperature further reveals that the Lokop hot spring is characterized by temperatures ranging from 36 to 38.5 °C, with the lowest recorded in 2023 and the highest in 2020, indicating an increased heat supply from the geothermal reservoir. We believe that the results of this research offer essential insights for enhanced characterisation of the geothermal system in the Lokop field.
Most geological maps are preserved as scanned or archived images, impeding direct access to vector data. Existing vectorization methods for such maps largely focus on region segmentation and remain insufficient for fine semantic recognition and vectorization of key linear features, such as faults and geological boundaries. To address this, we propose an intelligent linear-feature vectorization approach based on GWSU-Net (Geological-map Weighted Semi-supervised U-Net). This method builds upon the U-Net semantic segmentation framework and integrates pixel-level weighting, a comprehensive confidence evaluation, and a human-in-the-loop semi-supervised iterative strategy to automatically identify and vectorize faults and geological boundaries. We constructed an experimental dataset from 19 archived geological maps, yielding 32,582 image patches of 128 × 128 pixels. Three maps were used for initial training, ten for iterative augmentation, three as a validation set for model selection, and three as an independent test set. All final metrics are reported exclusively on the independent test maps, which were excluded from model training, pseudo-label generation, manual correction, and sample re-injection. On the independent test set, the final model achieved F1 scores of 0.8634 for faults and 0.8504 for geological boundaries. Vector-level evaluation yielded a line-length completeness of 0.8657 and a mean offset distance of 1.5008 pixels. These results demonstrate that the proposed method can stably extract linear features from archived geological maps with limited manual annotation.
The Weather Research and Forecasting (WRF) Model is widely used for numerical weather prediction, research, operational forecasting, and education. That flexibility is valuable, but it also makes initial configuration a common source of preventable errors. Before model execution begins, users must translate a scientific objective into domain geometry, map-projection settings, static-data paths, timing controls, physics and dynamics selections, and, in some workflows, data-assimilation and observation-processing configurations. WRF Configuration Studio addresses this problem with a browser-based environment for WRF domain design, WRF Preprocessing System (WPS) namelist editing, complete-catalog namelist.input configuration, WRF Data Assimilation (WRFDA) group visibility, observation-processing (OBSPROC) export as namelist.obsproc, Registry-derived metadata review, and non-blocking compatibility guidance. The domain workflow also provides dedicated Northern and Southern Hemisphere polar map views for high-latitude configuration, including hemisphere-aware Polar Stereographic routing and projection-specific geographic reference overlays. This project continues a practical lineage from early National Oceanic and Atmospheric Administration/Forecast Systems Laboratory graphical WRF Standard Initialization workflows and WRF Domain Wizard while expanding the scope from domain generation into a broader configuration environment. This manuscript presents the motivation, design, implementation, workflow, metadata architecture, software capabilities, limitations, and development path of WRF Configuration Studio version 1.0.0. The software is intended to improve configuration transparency and reduce preventable structural errors while leaving model execution, scientific configuration choices, and forecast validation to the official WRF-family tools and the user.
Underwater image enhancement is a challenging ill-posed problem due to the absorption and scattering of natural light in water. Nowadays, most existing underwater image enhancement methods based on convolutional neural networks are committed to boost the network performance by increasing the depth or width of the network, which not only greatly increases the computational load, but also ignores the interaction of features at different depths. To address these problems, a content-guided attention-based feature fusion network for underwater image enhancement is proposed to boost the performance of network. Specifically, we develop a multiscale feature extraction module and a cascading attention-aware enhancement module to extract global and local features of the network respectively. In addition, we construct a feature fusion module based on content-based guided attention, which regulates the features by learning the spatial weights to solve the mismatching of receptive field during the fusion process, and realizes the full interactive fusion between the global and local features of the network. Comprehensive evaluations on three benchmark datasets demonstrate that the proposed method achieves strong and competitive performance in both qualitative comparisons and quantitative evaluations.
The growing demand for rare earth elements (REEs) necessitates efficient exploration methods for ion-adsorption deposits, which are economically vital but costly to evaluate using traditional trace-element analysis. While machine learning (ML) offers promising predictive capabilities, its “black-box” nature often limits trust and practical application in geoscience. This study addresses this gap by introducing the Explainable Boosted Geochemical Trees (EBGT) framework, a supervised learning approach that embeds domain knowledge directly into model training. By incorporating monotonic constraints based on established geochemical principles (e.g., enforcing positive relationships for Al₂O₃ and negative for SiO₂) and engineering diagnostic oxide ratios (SiO₂/Al₂O₃, K₂O/Na₂O, Fe₂O₃/TiO₂), EBGT ensures predictions align with the processes controlling REE adsorption. Applied to major oxide data from weathered granites in Malaysia, EBGT was benchmarked against logistic regression, Random Forest, SVM, and k-NN. The results show that EBGT achieves top-tier predictive performance, with perfect recall, an AUC-ROC of 0.98, and superior probability calibration. More importantly, SHAP interpretability analysis confirms that geochemically meaningful features drive the model’s decisions, with K₂O/Na₂O emerging as the most influential predictor. This work demonstrates that embedding geological reasoning into ML through monotonic constraints yields models that are not only accurate but also transparent and actionable. EBGT thus provides a cost-effective, interpretable tool for early-stage REE prospecting, reducing reliance on expensive assays while offering geologically coherent insights for exploration targeting.
Systematic mineral prospectivity mapping (MPM) remains limited in the Beposo Area of Ghana’s Birimian Supergroup, despite the well-documented orogenic gold endowment of the wider Birimian terrane. Following a mineral-systems approach, the documented structural and hydrothermal controls on gold in the area were translated into explicit, testable expectations against which the model outputs were subsequently assessed. Random forest (RF) and decision tree (DT) classifiers were trained on eight geoscientific predictors derived exclusively from airborne magnetic and radiometric datasets; geochemical data were used solely to label the 67 known gold occurrences. Hyperparameters were optimised by 10-fold GridSearchCV within a spatially independent training partition, and both models were evaluated on the identical held-out test set using ROC and precision–recall curves, classification metrics and the Brier score. The RF-based MPM achieved an AUC of 0.93, F1-score of 0.85 and Brier score of 0.134, compared with 0.85, 0.81 and 0.152 respectively for the DT-based MPM. Because the held-out set is small, the bootstrap confidence intervals of the two models overlap; the RF is therefore reported as retaining the higher central estimate on every metric rather than as significantly superior. Lineament density ranked as the most influential single predictor in both models, although the structural-magnetic and hydrothermal-alteration predictor groups contributed comparably overall. The high-prospectivity class of the RF-based MPM captured 92.5
Graph-based models are increasingly used to represent spatial dependence in environmental systems, but their apparent value may be overstated when the target variable exhibits strong temporal persistence. We developed a lag-controlled residual benchmark to test whether graph connectivity adds predictive information beyond a selected lag-1 baseline. Annual county-level net primary productivity (NPP) was reconstructed from MOD17A3HGF.061 for 382 counties in the Yellow River Basin during 2003–2022. Lag-1 NPP was retained as the baseline, while Euclidean-Static, Hydro-Static, Hydro-Dynamic, and identity-skip graph models were trained to predict only the residual not captured by that baseline. Across 20 random seeds and an independent 2020–2022 test period, the lag-1 baseline achieved the lowest RMSE (0.0404; R² = 0.8977), whereas the graph models produced RMSE values of 0.0497–0.0506. The Hydro-Dynamic graph did not significantly outperform the symmetric Hydro-Static graph (paired Wilcoxon test, p = 0.0696), and no dry, normal, wet, or subbasin group showed a robust dynamic advantage. Conventional residual-target spatial models also remained weaker than the lag-1 baseline, with geographically weighted regression performing best among them (RMSE = 0.0443). These findings indicate that, at the annual county scale, increasingly complex spatial connectivity did not provide stable predictive gains beyond the selected lag-1 benchmark. The proposed framework provides a reproducible diagnostic for distinguishing temporal baseline skill from incremental graph-based information in geospatial environmental modelling.
Groundwater quality plays a vital role in water resource management, influencing human health, ecosystems, and economic development. This study integrates physical and chemical parameters with machine learning to model groundwater quality both temporally and spatially, enhancing decision-making in agriculture and environmental planning. Using the Water Quality Index (WQI), a framework was developed for groundwater quality zoning and prediction. The study focused on Lake Urmia’s eastern basins, analyzing data from 54 stations between 2008 and 2021. Annual WQI values were calculated based on chemical parameters and mapped using inverse distance weighting to visualize spatial and temporal variations. Three predictive models were employed to forecast WQI for the subsequent year: Artificial Neural Networks (ANN), a hybrid Convolutional Neural Network with Long Short-Term Memory (CNN-LSTM), and a hybrid model combining Successive Variational Mode Decomposition (SVMD) with CNN-LSTM (SVMD-CNN-LSTM). SVMD is a signal decomposition technique that separates input data into distinct intrinsic modes to reduce non-stationarity and noise, thereby improving forecasting accuracy. Results showed a deterioration in groundwater quality in Lake Urmia’s northeastern region in 2021, with WQI values falling into the very poor category. Model performance evaluation revealed that CNN-LSTM outperformed ANN, achieving a Root Mean Square Error (RMSE) of 46.881, a coefficient of determination (R²) of 0.921, a Nash–Sutcliffe Efficiency (NSE) of 0.791, Mean Absolute Percentage Error (MAPE) of 0.097. The SVMD-CNN-LSTM model further improved predictive accuracy, yielding an RMSE of 29.654, R² of 0.967, NSE of 0.946, and MAPE of 0.032. The SVMD-CNN-LSTM model reduced the error by 58.28
In supervised mineral prospectivity mapping (MPM), the selection of non-deposit samples continues to be one of the main challenges and an important source of uncertainty. In this study, a framework integrating geological constraint with point pattern analysis is proposed for the selection of non-deposit locations. Based on this strategy, ten different negative datasets were generated and used to construct ten mineral prospectivity models using a gray wolf optimized random forest (GWO-RF) algorithm. The results show that: (1) many of the high-potential zones predicted by the GWO-RF show clear spatial association with known mineral occurrences (KMOs) and geological features favorable for hydrothermal copper mineralization; (2) the spatial patterns of the prospectivity models vary depending on the location of negative points; (3) feature importance analysis reveals that proximity to intrusive rocks and geochemical variables of Cu and Mo are the most influential predictors across all models. It is worth noting that intrusive rocks were also incorporated as a key geological constraint during the non-deposit sample selection stage; (4) although variations in feature importance are observed among the models due to changes in the spatial location of non-deposit samples, the ranking of the most influential variables remains generally stable; (5) the normalized density (ND) values of the models range from 4.88 to 13.28, with an average of 8.018, indicating that model performance is sensitive to the spatial distribution of non-deposit samples; and (6) the confidence index (CI) serves as a robust ensemble approach to integrate the outputs of the ten models, identifying several new high-confidence exploration targets with lower uncertainty. These targets are commonly located on or around intrusive rocks, associated volcanic units, and along faults, faults intersection, and the intersection of faults and intrusive rocks.
Hydrological drought forecasting is essential for water resource management in arid transboundary basins. This study compares a Transformer-enhanced Physics-Informed Neural Network (PhysicsSolver) with Support Vector Regression-Response Surface Method (SVR-RSM) for predicting drought indices (SRI, SSI, NSSI) at 1-, 3-, 6-, and 12-month scales using Helmand River streamflow data (1961–2014). Moving averages of streamflow and runoff served as model inputs. For standardized indices (SRI, SSI), both models achieved excellent performance (r = 0.95-1.00, NSE = 0.88–0.99) with optimal results at 6- and 12-month scales, suitable for seasonal monitoring. PhysicsSolver showed marginal improvements over SVR-RSM. However, non-stationary NSSI forecasting proved challenging for both approaches. PhysicsSolver demonstrated advantages at 1-month scale (NSE = 0.98 vs. 0.88), but both models failed at 3–6 months with negative NSE values. Results indicate that while physics-informed architectures improve short-term non-stationary prediction, substantial methodological advances are needed for medium-term forecasting in transboundary basins where anthropogenic factors drive non-stationarity.
Ship detection in coastal Synthetic Aperture Radar (SAR) imagery remains challenging due to severe semantic ambiguity between ships and coastal infrastructure, together with environmental and acquisition variations that produce unstable feature representations and unreliable detection in cluttered near-shore regions. Existing approaches predominantly rely on implicit feature learning or feature-level invariance, without explicitly modeling cross-condition variability. To address these limitations, this paper proposes a Cross-Condition Distributional Stability Learning (CSDL) framework for robust SAR ship detection. The proposed method reformulates detection as a distribution-constrained learning problem, where each spatial region is represented as a condition-dependent latent feature distribution characterized by its mean and covariance. The novelty of the proposed framework lies in the learning formulation that unifies established mathematical tools under a distributional stability objective, rather than in the individual techniques themselves. Instead of enforcing feature-level invariance or aligning predefined domains, CSDL learns stability as a bounded cross-condition variability property of spatially localized feature distributions. A Wasserstein distance-based objective regularizes cross-condition feature dispersion relative to an aggregated reference distribution, while an uncertainty-driven relative stability learning mechanism uses the covariance trace to quantify feature variability and distinguish reliable ship regions from unstable background clutter. The framework is further supported by SAR imaging physics, where consistent ship scattering and stochastic sea clutter provide a physical interpretation of distributional stability. Experimental results on benchmark datasets demonstrate improved detection accuracy, robustness, and generalization under varying environmental conditions. The proposed framework is governed by a unified stability-oriented learning objective, in which Gaussian distribution modeling, Wasserstein-based feature dispersion regularization, uncertainty-driven ranking, and stability-aware modulation are jointly integrated to optimize cross-condition distributional stability.
Earthquakes are natural hazards that pose significant risks to social, economic, and infrastructural systems. Iran, located within the seismically active Alpine-Himalayan orogenic belt, frequently experiences damaging seismic events. This study presents a modified probabilistic seismic hazard analysis (PSHA) to assess seismic hazards in the Nowshahr Port region of northern Iran. An extensive earthquake catalog was compiled, incorporating 21 historical and 1115 instrumental events with moment magnitudes (Mw) ≥ 3. Fourteen potential seismic sources were identified and modeled using geophysical, tectonic, geological, and seismic data. Seismicity parameters were estimated via Kijko’s method, and the spatial hazard contribution of each source was analyzed across various magnitude ranges. Peak ground acceleration (PGA) on rock outcrop was computed using the OpenQuake engine, incorporating a regionally appropriate ground motion prediction equation (GMPE) within the framework of the modified PSHA. The study area was discretized into a 5 km grid (7061 points), then PGA and spectral acceleration (SA) on bedrock, with 5
Accurate prediction of land surface temperature (LST) is essential for understanding regional thermal environments, yet seasonal LST modelling remains challenging in temperate continental regions where summer and winter surface processes differ strongly. This study developed an interpretable machine learning framework to predict seasonal MODIS clear-sky daytime LST across Rostov Oblast, Russia, using multi-source Earth observation data from 2015 to 2024. Seasonal raster stacks were generated for summer and winter using MODIS LST, spectral indices, aerosol optical depth, albedo, evapotranspiration, precipitation, elevation, population density, and spatial control terms. Season-specific feature selection was performed using the 2015–2020 training period and was checked with variance inflation factor diagnostics. Model transferability was evaluated using random split validation, spatial block cross-validation, expanding temporal validation during 2021–2023, an independent 2024 test, spatial block bootstrap uncertainty analysis, and cross-product consistency assessment. Five regression models were compared: multiple linear regression (MLR), random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and deep neural network (DNN). The final summer LightGBM model achieved strong independent performance in 2024, with R2 = 0.848, RMSE = 1.132 °C, MAE = 0.909 °C, and Lin’s CCC = 0.930. The winter DNN model showed lower but acceptable accuracy, with R2 = 0.705, RMSE = 1.906 °C, MAE = 1.506 °C, and Lin’s CCC = 0.766. Residual Moran’s I indicated remaining spatial clustering, especially in winter, suggesting that some spatially structured cold-season processes remained unmodelled. SHapley Additive exPlanations (SHAP) and partial dependence plots (PDP) indicated season-specific model-based associations: summer predictions were primarily structured by NDBI, whereas winter predictions depended more strongly on spatial control terms, albedo, and precipitation. The framework provides a spatially explicit and interpretable approach for seasonal clear-sky daytime LST prediction in understudied temperate continental landscapes.
Radiogenic heat production (RHP) is important in understanding the thermal development of the continental crust and evaluating geothermal resources. This study analyzed compiled radiometric data from 92 rock samples collected in four areas of southwest Nigeria—Odeda, Lagelu, Okpella, and Ilesa—using a combination of statistical, geophysical, and computational approaches, including an initial 1-D heat flow (HF) model, to assess the region’s geothermal potential, a dimension of the area’s basement geology that has received little attention to date. Because reliable geothermal baselines are still scarce for Nigeria’s crystalline basement, the results are intended to support the country’s search for indigenous, low-carbon energy sources and to guide future site-selection for exploration. The principal radioactive elements have average concentrations of 1.32 ± 1.02
Lithium is a critical element for today’s energy transition, as its low atomic mass and high electrochemical potential make it ideal for rechargeable lithium-ion batteries used in electric vehicles, grid-scale storage, and portable electronics. Formed through diverse geological processes, lithium has recently been recognized in geothermal fluids as a cleaner alternative to conventional sources. Because lithium has historically been underrepresented in routine geothermal fluid analyses, low-cost predictive screening from existing hydrogeochemical measurements can help identify promising fluids for further characterization. Despite the growing use of machine learning in geothermal prediction studies, no previous work has applied leakage-aware cross-validation with calibrated uncertainty quantification to dissolved lithium estimation from compiled, multi-source geothermal hydrogeochemistry. The primary contribution is the domain application itself: a curated 86-samples Turkish geothermal dataset, a reproducible and leakage-controlled evaluation protocol, and an empirical assessment of what routine chemistry can and cannot resolve about dissolved lithium. Four models (a compact neural network, Random Forest, XGBoost, and Ridge regression) were evaluated on a dataset of 86 thermal springs and wells compiled from published sources across Türkiye, predominantly from Western Anatolia, split into 74 training and 12 held-out test samples. A five-fold GroupKFold cross-validation over 26 location groups prevented spatial leakage. On the 12 samples test set, drawn entirely from Western Anatolian and Marmara fields already represented in training, the neural network achieved the lowest error (RMSE = 1.18 mg/L, R² = 0.95), closely followed by Ridge (R² = 0.91); both metrics are sensitive to a single high-Li end-member (Tuzla, Li = 22.6 mg/L). A paired bootstrap test showed the difference on this test set is not statistically significant (p = 0.42). Under leave-one-group-out cross-validation across all 26 locations, however, the neural network (R² = 0.67) substantially outperformed Ridge (R² = −226), which failed to extrapolate to held-out high-Li fields, indicating that the nonlinear model generalizes more robustly across geochemical domains. Prediction intervals calibrated from cross-validation residuals achieved 91.7–100
Natural disaster recognition is crucial in disaster management, emergency response, and damage assessment. In modern scenarios, even leveraging Deep Learning (DL), specifically CNNs and hybrid learning models, delivers significant solutions, there are still many challenges in the field, such as redundant feature extraction, high computational cost, and decreased reliability in noisy and heterogeneous environments. These constraints may affect their utility in disaster contexts, where rapid verdict-creation is acute. To tackle such challenges, this work proposes a novel multi-level feature-learning framework that incorporates adaptive image enhancement, hybrid segmentation, multi-texture feature extraction, attention-based classification, and a self-adaptive optimization strategy within a single architecture for image classification of natural disasters. The proposed system combines the Adaptive Long Recursive Kalman Filter (ALRKF) and a Bi-LSTM (Bidirectional Long Short-Term Memory) network to reduce noise and enhance images across different environments adaptively. The hybrid U-ResSegNet architecture combines the best of both architectures, U-Net and ResNet-50, to achieve better segmentation results, enabling the extraction of regions without losing fine spatial details or high-level semantic information. The segmented images are then described using Local Binary Pattern (LBP) to capture complementary local texture characteristics, and Local Ternary Pattern (LTP) and GLCM (Grey-Level Co-occurrence Matrix) to capture complementary global texture characteristics. A feature classification-based attentive mechanism network is leveraged to classify the resulting feature representations, and the parameters are optimized using the proposed Self-Adaptive Driving Training Optimization (SA-DTO) algorithm, which enhances the convergence stability and learning efficiency. The Multi-Class Disaster Images Dataset and the Disaster Images Dataset (CNN-Model) tested the proposed framework, achieving classification accuracies of 98.97
