Artificial Intelligence in Geosciences最新文献

In exploration geochemistry, advances in the detection limit, breadth of elements analyze-able, accuracy and precision of analytical instruments have motivated the re-analysis of legacy samples to improve confidence in geochemical data and gain more insights into potentially mineralized areas. While a re-analysis campaign in a geochemical exploration program modernizes legacy geochemical data by providing more trustworthy and higher-dimensional geochemical data, especially where modern data is considerably different than legacy data, it is an expensive exercise. The risk associated with modernizing such legacy data lies within its uncertainty in return (e.g., the possibility of new discoveries, in primarily greenfield settings). Without any advanced knowledge of yet unanalyzed elements, the importance of re-analyses remains ambiguous. To address this uncertainty, we apply machine learning to multivariate geochemical data from different regions in Canada (i.e., the Churchill Province and the Trans-Hudson Orogen) in order to use legacy geochemical data to predict modern and higher dimensional multi-elemental concentrations ahead of planned re-analyses. Our study demonstrates that legacy and modern geochemical data can be repurposed to predict yet unanalyzed elements that will be realized from re-analyses and in a manner that significantly reduces the latency to downstream usage of modern geochemical data (e.g., prospectivity mapping). Findings from this study serve as a pillar of a framework for exploration geologists to predictively explore and prioritize potentially mineralized districts for further prospects in a timely manner before employing more invasive and expensive techniques.

Strong convective systems and the associated heavy rainfall events can trig-ger floods and landslides with severe detrimental consequences. These events have a high spatio-temporal variability, being difficult to predict by standard meteorological numerical models. This work proposes the M5Images method for performing the very short-term prediction (nowcasting) of heavy convective rainfall using weather radar data by means of a convolutional recurrent neural network. The recurrent part of it is a Long Short-Term Memory (LSTM) neural network. Prediction tests were performed for the city and surroundings of Campinas, located in the Southeastern Brazil. The convolutional recurrent neural network was trained using time series of rainfall rate images derived from weather radar data for a selected set of heavy rainfall events. The attained pre-diction performance was better than that given by the persistence forecasting method for different prediction times.

Data-driven prediction of time series is significant in many scientific research fields such as global climate change and weather forecast. For global monthly mean temperature series, considering the strong potential of deep neural network for extracting data features, this paper proposes a data-driven model, ResGraphNet, which improves the prediction accuracy of time series by an embedded residual module in GraphSAGE layers. The experimental results of a global mean temperature dataset, HadCRUT5, show that compared with 11 traditional prediction technologies, the proposed ResGraphNet obtains the best accuracy. The error indicator predicted by the proposed ResGraphNet is smaller than that of the other 11 prediction models. Furthermore, the performance on seven temperature datasets shows the excellent generalization of the ResGraphNet. Finally, based on our proposed ResGraphNet, the predicted 2022 annual anomaly of global temperature is 0.74722 °C, which provides confidence for limiting warming to 1.5 °C above pre-industrial levels.

The aim of the current work is to compare susceptibility maps of landslides produced using machine learning techniques i.e. multilayer perception neural nets (MLP), kernel logistic regression (KLR), random forest (RF), and multivariate adaptive regression splines (MARS); novel ensemble approaches i.e. MLP-Bagging, KLR-Bagging, RF-Bagging and MARS-Bagging in the Kurseong-Himalayan region. For the ensemble models the RF, KLR, MLP and MARS were used as base classifiers, and Bagging was used as meta classifier. Another objective of the current work is to introduce and evaluate the effectiveness of the novel KLR-Bagging and MARS-Bagging ensembles in susceptibility to landslide. Compiling 303 landslide locations to calibrate and test the models, an inventory map was created. Eighteen LCFs were chosen using the Relief-F and multi-collinearity tests for mapping the landslide susceptibility. Applying receiver operating characteristic (ROC), precision, accuracy, incorrectly categorized proportion, mean-absolute-error (MAE), and root-mean-square-error (RMSE), the LSMs were subsequently verified. The different validation results showed RF-Bagging (AUC training 88.69% & testing 92.28%) with ensemble Meta classifier gives better performance than the MLP, KLR, RF, MARS, MLP-Bagging, KLR-Bagging, and MARS-Bagging based LSMs. RF model showed that the slope, altitude, rainfall, and geomorphology played the most vital role in landslide occurrence comparing the other LCFs. These results will help to reduce the losses caused by the landslides in the Kurseong and in other areas where geo-environmental and geological conditions more or less similar.

Most known mineral deposits were discovered by accident using expensive, time-consuming, and knowledge-based methods such as stream sediment geochemical data, diamond drilling, reconnaissance geochemical and geophysical surveys, and/or remote sensing. Recent years have seen a decrease in the number of newly discovered mineral deposits and a rise in demand for critical raw materials, prompting exploration geologists to seek more efficient and inventive ways for processing various data types at different phases of mineral exploration. Remote sensing is one of the most sought-after tools for early-phase mineral prospecting because of its broad coverage and low cost. Remote sensing images from satellites are publicly available and can be utilised for lithological mapping and mineral exploitation. In this study, we extend an artificial intelligence-based, unsupervised anomaly detection method to identify iron deposit occurrence using Landsat-8 Operational Land Imager (OLI) satellite imagery and machine learning. The novelty in our method includes: (1) knowledge-guided and unsupervised anomaly detection that does not assume any specific anomaly signatures; (2) detection of anomalies occurs only in the variable domain; and (3) a choice of a range of machine learning algorithms to balance between explain-ability and performance. Our new unsupervised method detects anomalies through three successive stages, namely (a) stage I – acquisition of satellite imagery, data processing and selection of bands, (b) stage II – predictive modelling and anomaly detection, and (c) stage III – construction of anomaly maps and analysis. In this study, the new method was tested over the Assen iron deposit in the Transvaal Supergroup (South Africa). It detected both the known areas of the Assen iron deposit and additional deposit occurrence features around the Assen iron mine that were not known. To summarise the anomalies in the area, principal component analysis was used on the reconstruction errors across all modelled bands. Our method enhanced the Assen deposit as an anomaly and attenuated the background, including anthropogenic structural anomalies, which resulted in substantially improved visual contrast and delineation of the iron deposit relative to the background. The results demonstrate the robustness of the proposed unsupervised anomaly detection method, and it could be useful for the delineation of mineral exploration targets. In particular, the method will be useful in areas where no data labels exist regarding the existence or specific spectral signatures of anomalies, such as mineral deposits under greenfield exploration.

Gully erosion is one of the important problems creating barrier to agricultural development. The present research used the radial basis function neural network (RBFnn) and its ensemble with random sub-space (RSS) and rotation forest (RTF) ensemble Meta classifiers for the spatial mapping of gully erosion susceptibility (GES) in Hinglo river basin. 120 gullies were marked and grouped into four-fold. A total of 23 factors including topographical, hydrological, lithological, and soil physio-chemical properties were effectively used. GES maps were built by RBFnn, RSS-RBFnn, and RTF-RBFnn models. The very high susceptibility zone of RBFnn, RTF-RBFnn and RSS-RBFnn models covered 6.75%, 6.72% and 6.57% in Fold-1, 6.21%, 6.10% and 6.09% in Fold-2, 6.26%, 6.13% and 6.05% in Fold-3 and 7%, 6.975% and 6.42% in Fold-4 of the basin. Receiver operating characteristics (ROC) curve and statistical techniques such as mean-absolute-error (MAE), root-mean-absolute-error (RMSE) and relative gully density area (R-index) methods were used for evaluating the GES maps. The results of the ROC, MAE, RMSE and R-index methods showed that the models of susceptibility to gully erosion have excellent predictive efficiency. The simulation results based on machine learning are satisfactory and outstanding and could be used to forecast the areas vulnerable to gully erosion.

Compressional and shear sonic logs (DTC and DTS, respectively) are one of the effective means for determining petrophysical/geomechanical properties. However, the DTS log has limited availability mainly due to high acquisition costs. This study introduces a hybrid machine learning approach to generating synthetic DTS logs. Five wireline logs such as gamma ray (GR), density (RHOB), neutron porosity (NPHI), deep resistivity (Rt), and DTS logs are used as input data for three supervised-machine learning models including support vector machine for regression (SVR), deep neural network (DNN), and long short-term memory (LSTM). The hybrid machine learning model utilizes two additional techniques. First, as an unsupervised-learning approach, data clustering is integrated with general machine learning models for the purpose of improving model accuracy. All the machine learning models using the data-clustered approach show higher accuracies in predicting target (DTS) values, compared to non-clustered models. Second, particle swarm optimization (PSO) is combined with the models to determine optimal hyperparameters. The PSO algorithm proves time-effective, automated advantages as it gets feedback from previous computations so that is able to narrow down candidates for optimal hyperparameters. Compared to previous studies focusing on the performance comparison among machine learning algorithms, this study introduces an advanced approach to further improve the performance by integrating the unsupervised learning technique and PSO optimization with the general models. Based on this study result, we recommend the hybrid machine learning approach for improving the reliability and efficiency of synthetic log generation.

A separator and multiphase flow meters are considered the most accurate tools used to measure the surface oil flow rates. However, these tools are expensive and time consuming. Thus, this study aims to develop a correlation for accurate and quick evaluation of well surface flow rates and consequently the well inflow performance relationship. In order to achieve the abovementioned aim, this study uses artificial neural network (ANN) for flow rates prediction particularly in artificial lifted wells especially in the absence of wellhead pressure data. The ANN model is developed and validated by utilizing 350 data points collected from numerous fields located in Nile Delta and Western Desert of Egypt with inputs include; wellhead temperature, gas liquid ratio, water cut, surface and bottomhole temperatures, water cut, surface production rates, tubing cross section area, and well depth. The results of this study show that, the collected data are distributed as follows; 60% for training, 30% for testing and 10% for the validation processes with R² of 0.96 and mean square error (MSE) of 0.02. A comparison study is implemented between the new ANN correlation and other published correlations (Gilbert, Ros and Achong correlations) to show the robustness of the developed correlation. The results show that the developed correlation able to predict oil flow rates accurately with the lowest mean square error.

Attenuation of migration artifacts on Kirchhoff migrated seismic data can be challenging due to the relatively low amplitude of migration artifacts compared to reflections as well as the overlap in the kinematics of reflection and migration smiles. Several ‘conventional’ filtering methods exist and recently deep learning based workflows have been proposed. A deep learning workflow can be a simple and fast alternative to existing methods. In case of supervised training of a deep neural network using training data made by physics-based modelling or actual migrations is expensive and lacks diversity in terms of noise, amplitude, frequency content and wavelet. This can result in poor generalization beyond the training data without re-training and transfer learning. In this paper we demonstrate successful applications of migration smile separation using a conventional U-net architecture. The novelty in our approach is that we do not use synthetic data created from physics-based modelling, but instead use only synthetic data build form basic geometric shapes. Our domain of application is the migrated common offset domain, or simply the stack of the pre-stack migrated data, where reflections resemble local geology and migration smiles are upward convex hyperbolic patterns. Both patterns were randomly perturbed in many ways while maintaining their intrinsic features. This approach is inspired by the common practice of data augmentation in deep learning for machine vision applications. Since many of the standard data augmentation techniques lack a geophysical motivation, we have instead perturbed our synthetic training data in ways to make more sense for a signal processing perspective or given our ‘domain knowledge’ of the problem at hand. We did not have to retrain the network to produce good results on the field dataset. The large variety and diversity in examples enabled the trained neural network to show encouraging results on synthetic and field datasets that were not used in training.