Machine Learning in Geosciences
Use of the Convolutional Neural Network to invert electrical resistivity data, CNN-ERT 3D
CNN-ERT 3D:
M.T. VU et A. JARDANI, 2020, Convolutional neural networks with SegNet architecture applied to three-dimensional tomography of subsurface electrical resistivity. CNN-ERT, Geophysical Journal International.
Abstract: In general, the inverse problem of electrical resistivity tomography is treated using a deterministic algorithm to find a model of subsurface resistivity that can numerically match the apparent resistivity data acquired at the surface and has a smooth distribution that has been introduced as prior information. In this paper, we propose a new deep-learning approach for processing the 3D reconstruction of electrical resistivity tomography (ERT). This approach relies on the approximation of the inverse operator considered as a non-linear function linking the sections of apparent resistivity as input and the underground distribution of electrical resistivity as output. This approximation is performed with a large amount of known data to obtain an accurate generalization of the inverse operator by identifying during the learning process a set of parameters assigned to the neural networks. To train the network, the subsurface resistivity models are theoretically generated by a Geostatistical anisotropic Gaussian generator, and their corresponding resistivity apparent by solving numerically 3D Poisson equation. These data are formed in a way to have the same size and trained on the convolutional neural networks with Segnet architecture containing a 3-level encoder and decoder network ending with a regression layer. The encoders including the convolutional, max-pooling and nonlinear activator operations, are sequentially performed to extract the main features of input data in lower resolution maps. On the other side, the decoders are dedicated to upsampling operations in concatenating with feature maps transferred from encoders to compensate the loss of resolution. The tool has been successfully validated on different synthetic cases and with particular attention to how data quality in terms of resolution and noise affect the effectiveness of the approach.
Reconstruction of missing groundwater level data by using Long Short-Term Memory (LSTM) deep neural network
M. T. Vu , A. Jardani , N. Masse and M. Fournier , 2020: Reconstruction of missing groundwater level data by using Long Short-Term Memory (LSTM) deep neural network, Journal of hydrology
Abstract: Monitoring groundwater level (GWL) over long time periods is critical in understanding the variability of groundwater resources in the present context of global changes. However, in Normandy (France) for example, GWLs have only been systematically monitored for ~20 to 50 years. This study evaluates Long Short-Term Memory (LSTM) neural network modeling to reconstruct GWLs, fill gaps and extend existing time-series. The approach is illustrated by using available monitoring fluctuations in piezometers implanted in the chalk aquifer in the Normandy region, Northern France. Here GWL data recorded over 50 years at 31 piezometers in northwestern Normandy is employed to perform GWL prediction. To optimize the network performance, the most influential factors that impact the accuracy of prediction are first determined, such as the network architecture, data quantity and quality. The resulting network is adopted to reconstruct measurements in the piezometers step by step with an increment of missing observation time. The approach requires no calibration for the time-lag in data processing and the implementation relies only on the groundwater level fluctuations to retrieve missing data in the targeted piezometers.
Large-scale seasonal forecasts of river discharge by coupling local and global datasets with a stacked neural network: Case for the Loire River system
M.T. Vu, A. Jardani, M. Krimissa, F. Zaoui, N. Massei,Large-scale seasonal forecasts of river discharge by coupling local and global datasets with a stacked neural network: Case for the Loire River system, Science of The Total Environment, Volume 897,2023,165494,
Abstract
Accurate prediction of river discharge is critical for a wide range of sectors, from human activities to environmental hazard management, especially in the face of increasing demand for water resources and climate change. To address this need, a multivariate model that incorporates both local and global data sources, including river and piezometer gauges, sea level, and climate parameters. By employing phase shift analysis, the model optimizes correlations between the target discharge and 12 parameters related to hydrologic and climatic systems, all sampled daily. In addition, a stacked LSTM - a more complex neural network architecture - is used to improve information extraction ability.
Exploring river dynamics in the Loire-Bretagne basin and its surroundings, the investigation delves into predictions in daily time steps for one, three, and six months ahead. The resulting forecast features high accuracy and efficiency in predicting river discharge fluctuations, showcasing superior performance in forecasting drought periods over flood peaks. A detailed examination on data used highlights the significance of both local and global datasets in predicting river discharge, where the former dictates short-term predictions, while the latter drives long-range forecasts. Seasonally extended forecasting confirms a strong connection between the forecast leading time and the shift in data correlation, with lower correlation at a lag of 3 months due to seasonal changes affecting forecast quality, compensated by a higher correlation at a longer lag of 6 months. Such mutual effect in this multi-time-step forecasting improves the predictive quality of a six-month horizon, thus encourages progress in long-term prediction to a seasonal scale. The research establishes a practical foundation for effectively utilizing big data to leverage long-term forecasting of environmental dynamics.