2D probabilistic prediction of sparsely measured earth properties constrained by geophysical imaging fully accounting for tomographic reconstruction ambiguity
详细信息 查看全文
- 作者：Abduljabbar Asadi ; Peter Dietrich ; Hendrik Paasche
- 关键词：Geophysics ; Petrophysics ; Probabilistic prediction ; Tomography ; ANN ; Global search inversion
- 刊名：Environmental Earth Sciences
- 出版年：2016
- 出版时间：December 2016
- 年：2016
- 卷：75
- 期：23
- 全文大小：
- 刊物类别：Earth and Environmental Science
- 刊物主题：Geology; Hydrology/Water Resources; Geochemistry; Environmental Science and Engineering; Terrestrial Pollution; Biogeosciences;
- 出版者：Springer Berlin Heidelberg
- ISSN：1866-6299
- 卷排序：75
文摘
Many hydrological, environmental, or engineering exploration tasks require predicting spatially continuous scenarios of sparsely measured borehole logging data. We present a methodology to probabilistically predict such scenarios constrained by ill-posed geophysical tomography. Our approach allows for transducing tomographic reconstruction ambiguity into the probabilistic prediction of spatially continuous target parameter scenarios. It is even applicable to data sets where petrophysical relations in the survey area are non-unique, i.e., different facies related petrophysical relations may be present. We employ static two-layer artificial neural networks (ANNs) for prediction and additionally evaluate, whether the training performance of the ANNs can be used to rank geophysical tomograms, which are mathematically equal reconstructions of physical parameter distributions in the ground. We illustrate our methodology using a realistic synthetic database for maximal control about the prediction performance and ranking potential of the approach. For doing so, we try to link geophysical radar and seismic tomography as input parameters to porosity of the ground as target parameter of ANN. However, the approach is flexible and can cope with any combination of geophysical tomograms and hydrologic, environmental or engineering target parameters. Ranking of equivalent geophysical tomograms based on additional borehole logging data is found to be generally possible, but risks remain that the ranking based on the ANN training performance does not fully coincide with the closeness of geophysical tomograms to ground truth. Since geophysical field data sets do usually not offer control options similar to those used in our synthetic database, we do not recommend the utilization of recurrent ANNs to learn weights for the individual geophysical tomograms used in the prediction procedure.