Applied Computing and Geosciences最新文献

Detecting seismic events in real-time for prompt alerts and responses is a challenging task that requires accurately capturing P-wave arrivals. This task becomes even more challenging in regions like Indonesia, where widely spaced seismic stations exist. The wide station spacing makes associating the seismic signals with specific even more difficult. This paper proposes a novel deep learning-based model with three convolutional layers, enriched with dual attention mechanisms—Squeeze, Excitation, and Transformer Encoder (CNN-SE-T) —to refine feature extraction and improve detection sensitivity. We have integrated several post-processing techniques to further bolster the model's robustness against noise. We conducted comprehensive evaluations of our method using three diverse datasets: local earthquake data from East Java, the publicly available Seismic Waveform Data (STEAD), and a continuous waveform dataset spanning 12 h from multiple Indonesian seismic stations. The performance of the CNN-SE-T P-wave detection model yielded exceptionally high F1 scores of 99.10% for East Java, 92.64% for STEAD, and 80% for the 12-h continuous waveforms across Indonesia's network, demonstrating the model's effectiveness and potential for real-world application in earthquake early warning systems.

The flood hazards in the southwest coastal region of India in 2018 and 2020 resulted in numerous casualties and the displacement of over a million people from their homes. In order to mitigate the loss of life and resources caused by recurrent major and minor flood events, it is imperative to develop a comprehensive spatial flood zonation map of the entire area. Therefore, the main aim of the present study is to prepare a flood susceptible map of the southwest coastal region of India using synthetic-aperture radar (SAR) data and robust machine learning algorithms. Accurate flood and non-flood locations have been identified from the multi-temporal Sentinel-1 images. These flood locations are correlated with sixteen flood conditioning geo-environmental variables. The Boruta algorithm has been applied to determine the importance of each flood conditioning parameter. Six efficient machine learning models, namely support vector machine (SVM), k-nearest neighbors (KNN), artificial neural network (ANN), random forest (RF), partial least squares (PLS) and penalized discriminant analysis (PDA) have been applied to delineate the flood susceptible areas of the study region. The performance of the models has been evaluated using several statistical criteria, including area under curve (AUC), overall accuracy, specificity, sensitivity and kappa index. The results have revealed that all models have performed more than 90% of AUC due to the high precision of radar data. However, the RF and SVM models have outperformed other models in terms of all statistical parameters. The findings have identified approximately 13% of the study region as highly vulnerable to flood hazards, emphasizing the need for proper planning and management in these areas.

Accurate detection and characterization of seafloor morphologies are crucial for marine researchers and industries involved in underwater mapping, environmental monitoring, or resource exploration. Although their detection has relied on visual inspection of detailed bathymetries, few efforts to automate the process can be found in the literature. This study presents a novel MatLab computer code called POSIT (Feature Signature Detection) based on the convolution and correlation with a structural element containing the shape to search for. POSIT is successfully tested on both synthetic and real datasets, encompassing marine and terrestrial digital elevation models of different resolution and on a digital image. The centroids of submarine pockmarks and mounds, terrestrial volcanic craters and lunar craters are calculated with zero dispersion and perfect location, and their geometric parameters and confidence are provided.

Accurate segmentation of 3D micro CT scans is a key step in the process of analysis of the microstructure of porous materials. In polar ice core studies, the environmental effects on the firn column could be detected if the microstructure is digitized accurately. The most challenging task is to obtain the microstructure parameters of the bubbly ice section. To identify the minimum, necessary resolution, the bubbly ice micro CT scans with different resolutions (120, 60, 30, 12 $\mu m$) were compared object-wise via a region pairing algorithm. When the minimum resolution was found to be 60 $\mu m$, for generating the training/validation dataset, 4 ice core samples from a depth range of 96 to 108 meters (bubbly ice) were scanned with 120 $\mu m$ (input images) and another time with 4 times higher resolution (30 $\mu m$) to build ground truth. A specific pipeline was designed with non-rigid image registration to create an accurate ground truth from 4 times higher resolution scans. Then, two SOTA deep learning models (3D-Unet and FCN) were trained and later validated to perform super-resolution segmentation by taking input of $120\mu m$ resolution data and giving the output of binary segmented with two times higher resolution ($60\mu m$). Finally, the outputs of CNN models were compared with traditional rule-based and unsupervised methods on blind test data. It is observed the 3D-Unet can segment low-resolution scans with an accuracy of 96% and an f1-score of 80.8% while preserving microstructure having less than 2% error in porosity and SSA.

This study aims to tackle the obstacles linked with geological image segmentation by employing sophisticated deep learning techniques. Geological formations, characterized by diverse forms, sizes, textures, and colors, present a complex landscape for traditional image processing techniques. Drawing inspiration from recent advancements in image segmentation, particularly in medical imaging and object recognition, this research proposed a comprehensive methodology tailored to the specific requirements of geological image datasets. To establish the dataset, a minimum of 50 images per rock type was deemed essential, with the majority captured at the University of Las Palmas de Gran Canaria and during a field expedition to La Isla de La Palma, Spain. This dual-source approach ensures diversity in geological formations, enriching the dataset with a comprehensive range of visual characteristics. The study involves the identification of 19 distinct rock types, each documented with 50 samples, resulting in a comprehensive database containing 950 images. The methodology involves two crucial phases: initial preprocessing of the dataset, focusing on formatting and optimization, and subsequent application of deep learning models—ResNets, Inception V3, DenseNets, MobileNets V3, and EfficientNet V2 large. Preparing the dataset is crucial for improving both the quality and relevance, thereby to ensure the optimal performance of deep learning models, the dataset was preprocessed. Following this, transfer learning or more specifically fine-tuning is applied in the subsequent phase with ResNets, Inception V3, DenseNets, MobileNets V3, and EfficientNet V2 large, leveraging pre-trained models to enhance classification task performance. After fine-tuning eight deep learning models with optimal hyperparameters, including ResNet101, ResNet152, Inception-v3, DenseNet169, DenseNet201, MobileNet-v3-small, MobileNet-v3-large, and EfficientNet-v2-large, comprehensive evaluation revealed exceptional performance metrics. DenseNet201 and InceptionV3 attained the highest accuracy of 98.49% when tested on the original dataset, leading in precision, sensitivity, specificity, and F-score. Incorporating preprocessing steps further improved results, with all models exceeding 97.5% accuracy on the preprocessed dataset. In K-Fold cross-validation (k = 5), MobileNet V3 large excelled with the highest accuracy of 99.15%, followed by ResNet101 at 99.08%. Despite varying training times, models on the preprocessed dataset showed faster convergence without overfitting. Minimal misclassifications were observed, mainly among specific classes. Overall, the study's methodologies yielded remarkable results, surpassing 99% accuracy on the preprocessed dataset and in K-Fold cross-validation, affirming the efficacy in advancing rock type understanding.

Due to the nature of black-box machine learning (ML) models used in the spatial modelling field of environmental and natural hazards, the interpretation of predictive model outputs is necessary. For this purpose, we applied four interpretation techniques consisting of interaction plot, permutation feature importance (PFI) measure, shapley additive explanation (SHAP) decision plot, and accumulated local effects (ALE) plot to explain and interpret the output of an ML model applied to map land subsidence (LS) in the Nazdasht plain, Hormozgan province, southern Iran. We applied a stepwise regression (SR) algorithm and five ML models (Cforest (as a conditional random forest), generalized linear model (GLM), multivariate linear regression (MLR), partial least squares (PLS) and extreme gradient boosting (XGBoost)) to select important features and to map the LS hazard, respectively. Thereafter, several interpretation techniques were used to explain the spatial ML hazard model output. Our findings revealed that a GLM model was the most accurate approach to map LS in our study area, and that 24.3% of the total study area had a very high susceptibility to the LS hazard. According to the interpretation techniques, land use, elevation, groundwater level and vegetation were the most important variables controlling the LS hazard and also the most important variables contributing to the model’s output. Overall, human activities, especially the diversion of the route of one of the main tributaries feeding the plain and the recharging of groundwater five decades ago, intensified the current LS occurrence. Therefore, management activities such as water spreading projects upstream of the plain can be useful to mitigate LS occurrence in the plain.

Targeting reservoirs below seismic resolution presents a major challenge in reservoir characterization. High-resolution seismic data is critical for imaging the thin gas-bearing Khadro sand facies in several fields within the Lower Indus Basin (LIB). To truly characterize thin beds below tuning thickness, we showcase an optimally developed deep learning technique that can save up to 75% turn-around time while significantly reducing cost. Our workflow generates high-frequency acoustic impedance synthetics by utilizing a deep neural network (DNN) at the reservoir level vis-a-vis validating the results with existing geological facies. Simultaneously, we introduce continuous wavelet transform (CWT); wherein the three components (real, imaginary, and magnitude) are interrelated to obtain a resultant high-frequency seismic volume. A strong agreement is established at available wells to achieve a higher resolution seismic by injecting higher frequencies, which is then populated throughout the 3D cube. An excellent correlation is met with key seismic attributes extracted across the field for original and CWT-based synthetic seismic. The augmented seismic volume with enhanced frequency range substantiates the dominant frequency (F_d) and resolves thin beds, which is also validated with the help of wedge modeling of both acquired and high-frequency datasets. As a geologically valid solution, our approach effectively resolves an initially 54 m bed to ∼25 m. This deep-learning methodology is ideally suited to regions where the acquired seismic has limited resolution and lacks advanced reservoir characterization.

An analytical method to automatically characterize rock samples for geological or petrological purposes is here proposed, by applying machine learning approach (ML) as a protocol for saving experimental times and costs.

Proper machine learning algorithms, applied to automatically acquired microanalytical data (i.e., Electron Probe Micro Analysis, EPMA), carried out with a SEM-EDS microprobe on randomly selected areas from a petrographic polished thin section, are trained, used, tested, and reported.

Learning and Validation phases are developed with literature mineral databases of electron microprobe analyses on 15 main rock-forming mineral groups. The Prediction phase is tested using an eclogite rock from the Western Alps, considered as an unknown sample: randomly selected areas are acquired as backscattered images whose intervals of gray levels, appropriately set in the gray level histogram, allow the automated particle mineral separation: automated separating Oxford Instruments Aztec Feature ® packages and a mineral plotting software are applied for mineral particle separation, crystal chemical formula calculation and plotting.

Finally, a microanalytical analysis is performed on each separated mineral particle. The crystal chemical formula is calculated, and the final classification plots are automatically produced for any determined mineral. The final results show good accuracy and analytical ease and assess the proper nature of the unknown eclogite rock sample. Therefore, the proposed analytical protocol is especially recommended in those scenarios where a large flow of microanalytical data is automatically acquired and needs to be processed.

In the broad field of sensors and 3D information retrieval, bathymetric reconstruction from side-scan sonar imaging is associated with unique technical hurdles. Neural Networks have recently led to promising new solutions in this field, but the available methods tend to be complex and data-intensive in a way typically making their use in a riverine environment impossible. Throughout our work, we have focused on simplifying the problem-handling and treating compatibility with a riverine environment as priority. In our work, Long Short-Term Memory proved to be effective in a surprisingly simple form. Combined with traditional post-processing techniques in the GIS environment, like median filtered focal statistics, our workflow ultimately results in ∼0.259 m median of error on the evaluation dataset of the Dráva River.

This paper proposes two hybrid approaches based on Autoencoders (AEs) for long-term temperature prediction. The first algorithm comprises an AE trained to learn temperature patterns, which is then linked to a second AE, used to detect possible anomalies and provide a final temperature prediction. The second proposed approach involves training an AE and then using the resulting latent space as input of a neural network, which will provide the final prediction output. Both approaches are tested in long-term air temperature prediction in European cities: seven European locations where major heat waves occurred have been considered. The long-term temperature prediction for the entire year of the heatwave events has been analysed. Results show that the proposed approaches can obtain accurate long-term (up to 4 weeks) temperature prediction, improving Persistence and Climatology in the benchmark models compared. In heatwave periods, where the persistence of the temperature is extremely high, our approach beat the persistence operator in three locations and works similarly in the rest of the cases, showing the potential of this AE-based method for long-term temperature prediction.