3:48 PM - 4:05 PM
[MIS03-03] Artificial intelligence for facies classification based on high-resolution data from sediments of the Wadden Sea,Germany
★Invited Papers
Sediment facies classification is one of the first but crucial steps in the analysis of sediment records. It provides vital information, for example, for reconstructing past environmental variability and indicating locations of natural resources. The conventional way of facies classification is to macroscopically describe sediment structure and color and combine both with basic physical and chemical information - a time-consuming task heavily relying on the experience of the scientists in charge. Moreover, these sediment descriptions can not be re-evaluated quickly by further studies. In recent years, high-resolution and digitized elemental data are readily available from X-ray fluorescence core scanning techniques. Such data offer an opportunity for applying machine learning techniques to facilitate sediment facies classification. In 2016, an interdisciplinary research project, Wadden Sea Archive, was launched to investigate the Wadden Sea’s palaeo-landscapes and environments. The project has recovered more than 130 up to 5 m long sediment cores from the tidal flats, channels and off-shore around the island of Norderney (East Frisian Wadden Sea, Germany). These sediments were classified by the conventional method into 12 terrestrial and marine facies (e.g. glacioflucial sands, moraine, peat, tidal deposits, shoreface sediments). Altogether 92 cores were also scanned by an Itrax µ-XRF core scanner to obtain high-resolution (2000 µm) records of elemental data. This study gathered these data to develop and evaluate automatic facies classification models.
The raw elemental dataset was Centred log-ratio transformed to reduce bias caused by matrix effects and physical properties throughout the cores. Consequently, it was statistically represented by moving mean and standard deviation along each element to depict composite characteristics of a facies. The moving mean and standard deviation wrap the element intensities’ values and variations from the neighboring data points to a local point, respectively. The preprocessing (Principal component analysis) and classifier algorithms (Logistic regression, Support vector machine, Random forest) were implemented in the workflow to find the optimal solution. Our study provides confusion matrice, commonly used to describe details of prediction accuracy, and conjunction matrices, developed by this study to investigate the conjunctions between facies and visualize the models’ prediction characteristics. After cross-validation and evaluation of the test set, the model with the best performance was built by the Support vector machine algorithm without preprocessing, having 61 % accuracy in prediction. This is so far promising in comparison to a random guess from the 12 facies, which has a probability of 8 % of a correct prediction.
Our automated sediment facies classification based on the use of µ-XRF core scanning and machine learning techniques provides scientists with more resources to pursue additional research questions. The elemental data gives more possibilities for further research than the conventional method, usually qualitative description. We expect that our approach will contribute in developing a more comprehensive and time-efficient automatic sediment facies classification in the near future.
The raw elemental dataset was Centred log-ratio transformed to reduce bias caused by matrix effects and physical properties throughout the cores. Consequently, it was statistically represented by moving mean and standard deviation along each element to depict composite characteristics of a facies. The moving mean and standard deviation wrap the element intensities’ values and variations from the neighboring data points to a local point, respectively. The preprocessing (Principal component analysis) and classifier algorithms (Logistic regression, Support vector machine, Random forest) were implemented in the workflow to find the optimal solution. Our study provides confusion matrice, commonly used to describe details of prediction accuracy, and conjunction matrices, developed by this study to investigate the conjunctions between facies and visualize the models’ prediction characteristics. After cross-validation and evaluation of the test set, the model with the best performance was built by the Support vector machine algorithm without preprocessing, having 61 % accuracy in prediction. This is so far promising in comparison to a random guess from the 12 facies, which has a probability of 8 % of a correct prediction.
Our automated sediment facies classification based on the use of µ-XRF core scanning and machine learning techniques provides scientists with more resources to pursue additional research questions. The elemental data gives more possibilities for further research than the conventional method, usually qualitative description. We expect that our approach will contribute in developing a more comprehensive and time-efficient automatic sediment facies classification in the near future.