Murray Basin Cenozoic thickness

The Murray Basin is a saucer-shaped basin with flat-lying Cenozoic sediments up to approximately 600 m thickness (Brown and Stephenson, 1991). Constraints on the thickness of the Murray Basin have been compiled from: drillholes, reflection seismic profile interpretations, refraction seismic profiles and depth to magnetic basement estimates (Target_type.pdf). Target depths were sourced from Geoscience Australia, the national Groundwater Information System database (Http://www.bom.gov.au/water/groundwater/ngis/), the Geological Survey of Victoria (http://earthresources.vic.gov.au/earth-resources/geology-of-victoria/geological-survey-of-victoria) and the Geological Survey of South Australia (http://www.minerals.statedevelopment.sa.gov.au/geoscience/geological_survey). In addition, some of the magnetic depth estimates used data from McLean (2010). To constrain the thickness of Cenozoic cover where sediments were either absent or very thin we generated shallow-depth values in areas with post-Cenozoic geology and high topographic relief. In all, 5436 depth estimates were compiled (Target_depths.xlsx). The input datasets have been used to generate two predictive models of the thickness of Cenozoic sediments within the Murray Basin. The first model uses kriging of the depth estimates to generate a gridded surface using a local-area linear variogram model as a means of interpolating between constraints (Murray_Basin_kriging_Cenozoic_thickness.pdf; Murray_Basin_krig.tif -floating value grid). The second model uses a machine-learning approach where correlations between 17 supplementary datasets and 5436 depth estimates are used to derive a predictive model. We used a supervised learning algorithm known as Gaussian Process (GP) to generate the integrated predictive model. Gaussian Process is a non-parametric probabilistic approach to learning. It uses kernel functions to measure the similarity between points and predict values not seen from the training data (see Read_Me_GP.rtf). The supplementary datasets used in the model are listed in Table 1 and model variable settings can be found in read_me.rtf (Murray_Basin_GP_Cenozoic_thickness.pdf; Murray_Basin_GP_model.tif -floating value grid). Both approaches delineate the overall structure, geometry and thickness of the Murray Basin. The advantage of the machine learning approach is that it learns relationships between the depth and supplementary datasets which allow predictions in areas with limited constraints.ReferencesBrown, C. M. and Stephenson, A. E., 1991, Geology of the Murray Basin, southeastern Australia, Canberra, Bureau of Mineral Resources Bulletin 235, 430 p.McLean, M.A., 2010. Depth to Palaeozoic basement of the Gold Undercover region from borehole and magnetic data. GeoScience Victoria Gold Undercover Report 21. Department of Primary Industries, Victoria.Table 1. Supplementary input datasets used in predictive estimation of Murray Basin thickness, utilising a machine learning method Covariates* Description1 Latitude Gridded latitude values2 Longitude Gridded longitude values3 Elevation Terrain elevation – 90m shuttle DEM4 Distance from bedrock Euclidean distance from outcrop geology units older than Cenozoic5 Gravity Terrain and isostatic corrected Bouguer gravity6 Gravity 1228 Upward continued gravity at 1228 metres7 Gravity 2407 Upward continued gravity at 2407 metres8 Gravity 6605 Upward continued gravity at 6605 metres9 Gravity 18124 Upward continued gravity at 18124 metres10 Gravity 35524 Upward continued gravity at 35524 metres11 Gravity 49734 Upward continued gravity at 49734 metres12 Gravity 97479 Upward continued gravity at 97479 metres13 Gravity – 1k Isostatically corrected gravity subtracted from upward continued gravity at 1000 metres14 Magnetics 5km Upward continued magnetic anomaly grid at 5 km15 Magnetic 10km Upward continued magnetic anomaly grid at 10 km16 Magnetic 5-10km Upward continued 5km magnetic anomaly grid subtracted from upward continued 10 km magnetic anomaly grid17 Magnetic basement Depth to magnetic basement using the tilt method. *Primary datasets including gravity, magnetics and surface geology sourced from Geoscience Australia http://www.ga.gov.au/data-pubs/mapsElevation dataset used the 3 second (~90m) Shuttle Radar Topography Mission (SRTM) digital elevation model. https://pid.geoscience.gov.au/dataset/ga/72760.默里盆地（Murray Basin）为碟形盆地，发育产状平缓的新生代沉积物，最大厚度约600米（Brown与Stephenson, 1991）。 默里盆地厚度约束数据综合自钻孔、反射地震剖面解释、折射地震剖面及磁性基底埋深估算结果（详见Target_type.pdf）。目标深度数据源自澳大利亚地质调查局（Geoscience Australia）国家地下水信息系统数据库（http://www.bom.gov.au/water/groundwater/ngis/）、维多利亚州地质调查局（http://earthresources.vic.gov.au/earth-resources/geology-of-victoria/geological-survey-of-victoria）与南澳大利亚州地质调查局（http://www.minerals.statedevelopment.sa.gov.au/geoscience/geological_survey）。此外，部分磁性基底埋深估算数据取自McLean(2010)。 针对新生代沉积物缺失或极薄的区域，研究团队在新生代后地质发育且地形起伏较高的区域生成了补充性浅部埋深数据。本次研究共整合5436条深度估算数据（详见Target_depths.xlsx）。 研究基于输入数据集构建了两套默里盆地新生代沉积物厚度的预测模型。第一套模型采用克里金（kriging）插值方法，以局域线性变异函数模型为约束，在深度估算点间进行插值以生成网格化表面数据（Murray_Basin_kriging_Cenozoic_thickness.pdf；Murray_Basin_krig.tif——浮点值网格）。第二套模型采用机器学习方法，通过分析17项补充数据集与5436条深度估算数据间的相关性构建预测模型。研究使用了一种名为高斯过程（Gaussian Process, GP）的监督学习算法生成集成预测模型。高斯过程是一种非参数概率学习方法，通过核函数衡量样本点间的相似性，进而对训练集外的未知样本进行预测（详见Read_Me_GP.rtf）。模型所用补充数据集详见表1，模型变量设置可参考read_me.rtf（Murray_Basin_GP_Cenozoic_thickness.pdf；Murray_Basin_GP_model.tif——浮点值网格）。 两种建模方法均可刻画默里盆地的整体构造、几何形态与沉积物厚度。机器学习方法的优势在于可学习深度估算数据与补充数据集间的关联关系，从而在约束数据有限的区域实现厚度预测。 ### 参考文献 Brown, C. M. 与 Stephenson, A. E., 1991, 澳大利亚东南部默里盆地地质，堪培拉，矿产资源局公报235，共430页。 McLean, M.A., 2010. 基于钻孔与磁数据的金覆盖区古生界基底埋深。GeoScience Victoria金覆盖区报告21，维多利亚州初级产业部。 ### 表1 机器学习方法预测默里盆地沉积物厚度所用的补充输入数据集 |序号|协变量*|描述| |---|---|---| 1|纬度（Latitude）|网格化纬度值 2|经度（Longitude）|网格化经度值 3|高程（Elevation）|地形高程——90米分辨率航天飞机雷达地形测绘（Shuttle Radar Topography Mission, SRTM）数字高程模型 4|距基岩距离（Distance from bedrock）|距新生代之前出露地质单元的欧几里得距离 5|重力（Gravity）|经地形与均衡校正的布格重力（Bouguer gravity）异常 6|重力1228（Gravity 1228）|1228米高度延拓的重力异常 7|重力2407（Gravity 2407）|2407米高度延拓的重力异常 8|重力6605（Gravity 6605）|6605米高度延拓的重力异常 9|重力18124（Gravity 18124）|18124米高度延拓的重力异常 10|重力35524（Gravity 35524）|35524米高度延拓的重力异常 11|重力49734（Gravity 49734）|49734米高度延拓的重力异常 12|重力97479（Gravity 97479）|97479米高度延拓的重力异常 13|重力-1k（Gravity – 1k）|1000米高度延拓重力异常减去经均衡校正的重力异常 14|磁测5km（Magnetics 5km）|5千米高度延拓的磁异常网格 15|磁测10km（Magnetic 10km）|10千米高度延拓的磁异常网格 16|磁测5-10km（Magnetic 5-10km）|10千米高度延拓磁异常网格减去5千米高度延拓磁异常网格 17|磁性基底（Magnetic basement）|采用倾斜法计算的磁性基底埋深 *注：重力、磁测与地表地质等基础数据集源自澳大利亚地质调查局（Geoscience Australia）http://www.ga.gov.au/data-pubs/maps 本研究使用的高程数据采用3秒分辨率（约90米）的航天飞机雷达地形测绘任务（Shuttle Radar Topography Mission, SRTM）数字高程模型，数据链接：https://pid.geoscience.gov.au/dataset/ga/72760。