Earth Science Informatics最新文献

The elastic modulus of basalt is a significant engineering parameter required for many projects. Therefore, a total of 137 datasets of basalts from Digor-Kilittasi, Turkey, were used to predict the elastic modulus of intact rock (E_i) for this study. P wave velocity, S wave velocity, apparent porosity, and dry density parameters were employed as input parameters. In order to predict E_i, seven different models with two or three inputs were constructed, employing four different machine learning methods such as Support Vector Machine (SVM), Gaussian Process Regression (GPR), Ensembles of Tree (ET), and Regression Trees (RT). The performance of datasets, models, and methods was evaluated using the coefficient of determination (R²), Root Mean Squared Error (RMSE), Mean Squared Error (MSE), and Mean Absolute Error (MAE). This study presented and analyzed the performance of four machine learning methods. A ranking approach was employed to determine the best performing method and dataset. Based on these evaluations, all four machine learning techniques effectively estimate the value of E_i. While they can be used as an appropriate choice for estimating the elastic modulus of basaltic rocks, the ET approach appears to be the most successful method. However, the performance of the GPR is the worst according to model assessments. The average R² values for Model 1 through 7 of the ET method for the five test datasets are 0.97, 0.93, 0.89, 0.97, 0.91, 0.99, and 0.99, respectively. The the average R² values for GPR from Models 1 to 7 for the five test datasets are 0.73, 0.55, 0.69, 0.48, 0.47, 0.73, 0.56, respectively. An additional indication that the ET performed better than all the other methods was the Taylor diagram, which made it simple to determine how well the model predictions matched the observations. Furthermore, these findings validate the performance of the machine learning techniques employed in this study as valuable instruments for future investigations into the modeling of complex engineering issues. The results of this study suggest that machine learning algorithms can help reduce the need for high-quality core samples and labor-intensive procedures in predicting the elastic modulus of basaltic rocks, resulting in time and cost savings.

Underwater object detection is vital for diverse applications, from studies in marine biology to underwater robotics. However, underwater environments pose unique challenges, including reduced visibility due to color distortion, light attenuation, and complex backgrounds. Traditional computer vision methods have limitations, prompting the implementation of deep learning, for underwater object detection. Despite progress, challenges persist, such as visual degradation, scale variations, diverse marine species, and complex backgrounds. To address these issues, we propose Feature-Adaptive FPN with Multiscale Context Integration (FA-FPN-MCI), a novel deep-learning algorithm aimed at enhancing both detection and domain generalization performance. We integrate the Style Normalization and Restitution (SNR) module for domain generalization, Receptive Field Blocks (RFBs) for fine-grained detail capture, and a twin-branch Global Context Module (TBGCM) for multiscale context information. We enhance lateral connections within the Feature Pyramid Network (FPN) with deformable convolution. Experimental outcome reveal that the proposed method attains mean average precision of 84.2%. Additionally, other performance metrics were evaluated, and outperforming all other methods used for comparison.

Tropospheric radio refractivity is a significant atmospheric phenomenon that affects the propagation of radio signals, and can impact the design and operation of wireless communication systems. This study focuses on the development of an autoregressive model of tropospheric radio refractivity in Nigeria using artificial neural networks (ANNs). The proposed model utilizes atmospheric variables—temperature, pressure, and humidity—as inputs and predicts refractivity values with high accuracy. Descriptive statistics and data visualization techniques were used to gain insights into the relationships between the atmospheric variables and computed radio refractivity. It could be deduced from the results obtained that the developed ANN model accurately predicts tropospheric radio refractivity, with satisfactory performance indicators that include standard error (SE), root mean square error (RMSE), and correlation coefficient (R). It also demonstrates the reliability and robustness of the developed model, which could play an important role in improving the preparation and implementation routines of wireless communication systems. The study also identifies areas for further study, such as data availability, model complexity, and interpretability. Lastly, this work has further validated the suitability of applying ANNs to tropospheric radio refractivity model optimization, as it provides insights into the potential of the non-linear autoregressive modeling (NARX-ANN) approach for improving wireless communication systems.

Land subsidence, the loss of elevation of the earth's surface caused by natural and human-induced factors, has become a significant global concern. It poses substantial threats to urban planning, construction, and sustainable development. Monitoring and predicting regional land subsidence are particularly crucial. Interferometric Synthetic Aperture Radar (InSAR) and deep learning provide valuable insights into monitoring and predicting land subsidence. However, methods for accurate and long-term monitoring and predicting time series land subsidence still have limitations. Firstly, most models only utilize historical data and overlook the combined effects of various factors, including human activities and urbanization. Secondly, the spatiotemporal correlation of subsidence across different locations and times is underestimated. Thirdly, the nonlinearity of land subsidence is not adequately addressed. To address these challenges, this study assesses land deformation patterns from January 2018 to December 2022, using Sentinel-1 InSAR data processed through Small Baseline Subset-InSAR (SBAS-InSAR). The result shows that the annual average deformation rate ranged from -6.39 to 8.27 mm/year, with maximum cumulative subsidence and uplift of 27.62 mm and 36.62 mm, respectively. Subsequently, a GeoTemporal Transformer (GTformer) model based on the Transformer model is proposed. It captures nonlinearities and spatiotemporal correlations between land subsidence and influencing factors by generating spatiotemporal distance matrices. The results demonstrate the efficacy of the GTformer model in improving prediction accuracy by incorporating urbanization factors and constructing spatiotemporal distance matrices. Compared with traditional machine learning models, the R² of GTformer has increased by at least 14.6%, and compared with the standard Transformer, it has increased by 4%. The predictions closely align with observed subsidence patterns, highlighting the reliability. Moreover, this study underscores the critical role of urbanization factors in land subsidence mechanisms. The GTformer model provides a novel approach that integrates multiple factors and spatiotemporal correlation to predict land subsidence. The methodology offers a valuable tool for urban planners and decision-makers to effectively manage urban development and mitigate geological disaster risks.

Droughts are among the most hazardous and costly natural disasters and are hard to quantify and characterize. Accurate drought forecasting reduces droughts' devastating economic effects on ecosystems and people. Eastern Anatolia is the largest and coldest geographical region of Türkiye. Previous studies lack drought forecasting in the Eastern Anatolia (Upper Mesopotamia) Region, where agriculture is limited due to being under snow most of the year. This study focuses on the Euphrates basin, specifically the Tercan and the Tunceli meteorological stations of the Karasu River sub-basin, a vital Eastern Anatolia Region water resource. In this context, time series of 1-, 3-, 6-, 9-, and 12-month Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) values were created. The Tuned Q-factor Wavelet Transform (TQWT) method and Univariate Feature Ranking Using F-Tests (FSRFtest) were used for pre-processing and feature selection. Several models were created, such as stand-alone, hybrid, and tribrid. Machine Learning (ML) methods such as Artificial Neural Networks (ANN), Gaussian Process Regression (GPR), and Support Vector Machine (SVM) were conducted for the time series analyses. The GPR approach was concluded to perform better than the ANN and SVM at the Tercan station. In other words, GPR performs better in 80% of cases than SVM and ANN models. At the Tunceli station for the SPI output, SVM, which had a superior performance in 60% of the cases, demonstrated a performance comparable to GPR. At the same time, ANN once again exhibited an inferior performance. Similarly, for the SPEI output at the Tunceli station, no clear superiority was observed between the GPR and ANN methods. Because both methods were successful in 40% of cases. This study contributes by introducing a third concept to the stand-alone and hybrid model comparison of drought forecasting, adding tribrid models. It has been detected that the Hybrid and Tribrid ML methods lead to a 91% and 64% decrease relative root mean square error percentage compared stand-alone ML methods for SPEI and SPI in two stations. While the hybrid model at Tercan station was more successful in 80% of the cases, the hybrid model at Tercan station was more successful in 90% of the cases. While hybrid models were observed to be superior, tribrid models not only demonstrated performance close to the hybrid models but also provided advantages such as reducing computational load and shortening calculation time.

Hyperspectral image classification is crucial for a wide range of applications, including environmental monitoring, precision agriculture, and mining, due to its ability to capture detailed spectral information across numerous wavelengths. However, the high dimensionality and complex spatial-spectral relationships in hyperspectral data pose significant challenges. Deep learning, particularly Convolutional Neural Networks (CNNs), has shown remarkable success in automatically extracting relevant features from high-dimensional data, making them well-suited for handling the intricate spatial-spectral relationships in hyperspectral images.This study presents a hybrid approach for hyperspectral image classification, combining 3D Depthwise Separable Convolution (3D DSC) and Depthwise Squeeze-and-Excitation Network (DSENet). The 3D DSC efficiently captures spatial-spectral features, reducing computational complexity while preserving essential information. The DSENet further refines these features by applying channel-wise attention, enhancing the model's ability to focus on the most informative features. To assess the performance of the proposed hybrid model, extensive experimental studies were carried out on four commonly utilized HSI datasets, namely HyRANK-Loukia and WHU-Hi (including HongHu, HanChuan, and LongKou). As a result of the experimental studies, the HyRANK-Loukia achieved an accuracy of 90.9%, marking an 8.86% increase compared to its previous highest accuracy. Similarly, for the WHU-Hi datasets, HongHu achieved an accuracy of 97.49%, reflecting a 2.11% improvement over its previous highest accuracy; HanChuan achieved an accuracy of 97.49%, showing a 2.4% improvement; and LongKou achieved an accuracy of 99.79%, providing a 0.15% improvement compared to its previous highest accuracy. Comparative analysis highlights the superiority of the proposed model, emphasizing improved classification accuracy with lower computational costs.

Mineral grain recognition is an extremely important task in many fields, especially in mineral exploration, when trying to identify locations where precious minerals can possibly be found. The usual manual method would be to collect samples; a specialized individual using expensive equipment would manually identify and then count the grain minerals in the sample. This is a tedious task that is time-consuming and expensive. It is also limited because small portions of areas can be surveyed; even then, it might require extremely long periods. In addition, this process is still prone to human errors. Developing an automatic system to identify, recognize, and count grain minerals in samples from images would allow for more precise results than the time required by humans. In addition, such systems can be fitted on robots that collect samples, take images of the samples, and then proceed with the automated recognition and counting algorithm without human intervention. Vast amounts of land can be surveyed in this way. This paper proposes a modified approach for microscopic grain mineral recognition and classification using hybrid features and ensemble algorithms from images. The enhanced approach also included a modified segmentation approach, which enhanced the results. For 10 classes of microscopic mineral grains, using the modified approach and the ensemble algorithm resulted in an average accuracy of 84.01%. For 8 classes, the average reported accuracy is 94.93% using the Boosting ensemble learning with the C4.5 classifier. The results obtained outperform similar methods reported in the extant literature.

Effective drilling planning relies on understanding the rock mechanical properties, typically estimated from petrophysical data. Real-time estimation of these properties, especially static Young's modulus (({E}_{sta})), is crucial for geomechanical modeling, wellbore stability, and cost-effective decision-making. In this study, predictive models of ({E}_{sta}) were developed using mudlogging data from two vertically drilled wells (A and B) in the same field. ({E}_{sta}) was estimated from petrophysical data across the studied depth range in both wells using a field-specific equation. Outlier data were identified and removed by evaluating the cross plot of mechanical specific energy and drilling rate for Well A. The data from Well A were then randomly divided into training and testing sets. The algorithms, multi-layer perceptron neural networks, random forests, Gaussian process regression (GPR), and support vector regression, were adjusted and applied to the training data. The resulting models were evaluated on the test data. The GPR model demonstrated the lowest RMSE values in both the training (0.0075 GPa) and testing (0.4577 GPa) phases, indicating superior performance. To further assess the models, the overfitting index and scoring techniques were employed, revealing that the GPR model exhibited the lowest overfitting value and outperformed the other models. Consequently, the GPR model was selected as the best-performing model and was analyzed using Shapley additive explanation to evaluate the influence of each input feature on the output. This analysis indicated that depth had the greatest effect, while rotation speed had the least impact on the model's output. The application of the GPR model to predict ({E}_{sta}) in Well B demonstrated its high generalization capability. Therefore, it can be confidently stated that with additional data, this model could be effectively applied to similar depth ranges in other wells within the field. The study introduces innovations by applying GPR to predict ({E}_{sta}) from mudlogging data, addressing outlier impact on predictions, and developing a real-time ({E}_{sta}) prediction model for drilling.

In this study, classification and regression tree (CART) and multivariate adaptive regression spline (MARS) models are proposed to predict the stress concentration factor (SCF) around an extracted underground coal panel. Models are trained and tested using 120 collected datasets with 100 series allocated for models training and 20 datasets reserved for testing. For SCF prediction using the CART and MARS models, input parameters including overburden thickness (H), specific gravity of rock mass (γ), straight distance from the panel edge (D), and height of disturbed zone over the mined panel (H_d) are utilized, employing principal component analysis (PCA) to remove correlations. A predictive tree graph and 17 if–then rules with quantitative outputs are generated from the CART model, while a predictive equation is derived from the MARS technique for SCF prediction. The achieved values of the coefficient of determination (R²) for CART and MARS models are 0.940 and 0.957, respectively. Furthermore, obtained amounts of normalized root mean square error (NRMSE), variant account for (VAF), and performance index (PI) for CART are 0.043 92.473%, and 1.82, respectively. For the MARS model these values are 0.035, 95.419%, and 1.876,. Additionally, performance evaluations of the models using the Wilcoxon Signed Ranks and Friedman non-parametric tests, along with Taylor diagrams and error analysis demonstrate the reliability and suitability of the proposed models for SCF prediction. However, error and accuracy analyses confirm that MARS model yields more precise outputs, achieving 2.57% greater accuracy and 10.84% lower error than the CART model. Furthermore, the importance analysis demonstrated that both H and H_d have the highest importance on the SCF, while γ has the lowest, with importance values of 33.33% and 11.11%, respectively. Models verification based on the field SCF measurement confirms the models validity, as indicated by the relative errors of 6.83 for the MARS model and 7.05 for the CART model. Finally, a comparative analysis based on a case study data validates the practical application of the proposed models.

In recent years, technical accidents caused by natural disasters have caused huge losses. The purpose of this study is to develop a mathematical model to predict and prevent the risk of such accidents. The model applied machine learning to predict the risk of such accidents in the hope of providing risk visualization results for local governments. The expected impact of this research will benefit residents and public welfare organizations. In this study, Random Forest (RF), the K-Nearest Neighbor (KNN), the Back Propagation (BP) neural network, Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), and the Extreme Gradient Boosting (XGBoost) was applied to predict the risk value. At the same time, this study applied ArcGIS to spatially interpolate the risk prediction values to generate the risk map. The results demonstrated that the RF algorithm achieved the highest classification performance among the five algorithms tested. Specifically, the RF algorithm attained an accuracy of 0.874, an F1-Score of 0.887, and an Area Under the Curve (AUC) of 0.984. The three townships with the highest risk were Xueyan, Daibu, and Shanghuang, with the proportion of risk area accounting for 48.39%, 44.34% and 79.64% respectively. This study provides a reference for the local government, which can take targeted measures to prevent and control. For disaster managers, the risks for those high-risk areas should receive sufficient attention. The government should establish a real-time updated disaster database to monitor the development of the situation. Moreover, the development and acquisition of historical disaster data is worthy of encouragement.