High resolution conductivity mapping using regional AEM survey and machine learning.

The AEM method measures regolith and rocks' bulk subsurface electrical conductivity, typically to a depth of several hundred meters. AEM survey data is widely used in Australia for mineral exploration (i.e. mapping undercover and detection of mineralisation), groundwater assessment (i.e. hydro-stratigraphy and water quality) and natural resource management (i.e. salinity assessment). Geoscience Australia (GA) has flown Large regional AEM surveys over Northern Australia, including Queensland, Northern Territory and Western Australia. The surveys were flown nominally at 20-kilometre line spacing, using the airborne electromagnetic systems that have signed technical deeds of staging with GA to ensure they can be modelled quantitatively. Geoscience Australia commissioned the survey as part of the Exploring for the Future (EFTF) program. The EFTF program is led by Geoscience Australia (GA), in collaboration with the Geological Surveys of the Northern Territory, Queensland, South Australia and Western Australia, and is investigating the potential mineral, energy and groundwater resources in northern Australia and South Australia. We have used a machine learning modelling approach that establishes predictive relationships between the inverted flight-line modelled conductivity with a suite of national environmental and geological covariates. These covariates include terrain derivatives, gamma-ray radiometric, geological maps, climate derived surfaces and satellite imagery. Conductivity-depth values were derived from a single model using GA's deterministic 1D smooth-30-layer layered-earth-inversion algorithm. (Brodie and Richardson 2015). Three conductivity depth interval predictions are generated to interpolate the actual modelled conductivity data, which is 20km apart. These depth slices include a 0-50cm, 9-11m and 22-27m depth prediction. Each depth interval was modelled and individually optimised using the gradient boosted tree algorithm. The training cross-validation step used label clusters or groups to minimise over-fitting. Many hundreds of conductivity models are generated (i.e. ensemble modelling). Here we use the median of the models as the conductivity prediction and the upper and lower percentiles (95th and 5th) to measure model uncertainty. Grids show conductivity (S/m) in log 10 units. Reported out-of-sample r-squares for each interval in order of increasing depth are 0.74, 0.64, and 0.67. A decline in model performance with increasing depth was expected due to the decrease in suitable covariates at greater depths. Modelled conductivities seem to be consistent with the geological, regolith, geomorphological, and climate processes in the study area. The conductivity grids are at the resolution of the covariates, which have a nominal pixel size of 85 meters. Datasets in this data package include; 1. 0-50cm depth interval 0_50cm_median.tif; 0_50_upper.tif; 0_50_lower.tif 2. 9-11m depth interval 9_11m_median.tif; 9_11m_upper.tif; 9_11m_lower.tif 3. 22-27m depth interval 22_27_median.tif; 22_27_upper.tif; 22_27_lower.tif 4. Covariate shift; Cov_shift.tif (higher values = great shift in covariates) Reference: Ross C Brodie & Murray Richardson (2015) Open Source Software for 1D Airborne Electromagnetic Inversion, ASEG Extended Abstracts, 2015:1, 1-3, DOI: 10.1071/ ASEG2015ab197

航空电磁法（Airborne Electromagnetic, AEM）可测量风化层与岩石的整体地下体电导率，典型探测深度可达数百米。航空电磁测量数据在澳大利亚被广泛应用于矿产勘探（即覆盖区填图与矿化探测）、地下水评估（即水文地层学与水质监测）以及自然资源管理（即盐度评估）。
澳大利亚地质调查局（Geoscience Australia, GA）已在澳大利亚北部区域（包括昆士兰州、北领地与西澳大利亚州）开展了大型区域性航空电磁测量。该测量采用与该局签署技术实施协议的航空电磁系统，以确保可进行定量建模，飞行线间距标称值为20千米。
此项测量由澳大利亚地质调查局作为“未来勘探（Exploring for the Future, EFTF）”计划的一部分委托开展。该计划由澳大利亚地质调查局牵头，与北领地、昆士兰州、南澳大利亚州及西澳大利亚州的地质调查机构合作，旨在调查澳大利亚北部与南澳大利亚州的矿产、能源及地下水资源潜力。
本研究采用机器学习建模方法，建立反演得到的飞行线电导率模型与一系列全国性环境及地质协变量之间的预测关系。协变量涵盖地形衍生数据、伽马射线辐射测量数据、地质图、气候衍生表面数据以及卫星影像。
电导率-深度值通过澳大利亚地质调查局的确定性一维平滑30层层状地球反演算法（Brodie与Richardson, 2015）从单一模型中导出。
针对间距为20千米的实测电导率数据，我们生成了三个深度区间的预测结果以进行插值。这三个深度切片分别为0~50厘米、9~11米以及22~27米。每个深度区间均采用梯度提升树算法进行建模与单独优化。训练交叉验证步骤使用标签聚类或分组以最小化过拟合风险。
本研究生成了数百个电导率模型（即集成建模），最终采用模型中位数作为电导率预测结果，并以第95和第5百分位数来衡量模型不确定性。
栅格数据以10为底的对数单位显示电导率（单位：西门子/米，S/m）。按深度递增顺序，各深度区间的样本外决定系数（r-squared）分别为0.74、0.64和0.67。由于更深层的可用协变量减少，模型性能随深度增加而下降的情况符合预期。
模拟得到的电导率与研究区域内的地质、风化层、地貌及气候过程具有一致性。电导率栅格的分辨率与协变量一致，协变量的标称像素尺寸为85米。
本数据包包含以下数据集：
1. 0~50厘米深度区间：0_50cm_median.tif、0_50_upper.tif、0_50_lower.tif
2. 9~11米深度区间：9_11m_median.tif、9_11m_upper.tif、9_11m_lower.tif
3. 22~27米深度区间：22_27_median.tif、22_27_upper.tif、22_27_lower.tif
4. 协变量偏移：Cov_shift.tif（数值越高代表协变量偏移越大）
参考文献：Ross C Brodie & Murray Richardson (2015) 用于一维航空电磁反演的开源软件，ASEG扩展摘要，2015:1, 1-3，DOI: 10.1071/ASEG2015ab197