Abstract：
The radiological characterization of soil contaminated with natural radionuclides enables the classification of the area under investigation, the optimization of laboratory measurements, and informed decision-making on potential site remediation. Neural networks (NN) are emerging as a new candidate for performing these tasks as an alternative to conventional geostatistical tools such as Co-Kriging. This study demonstrates the implementation of a NN for estimating radiological values such as ambient dose equivalent (H*(10)), surface activity and activity concentrations of natural radionuclides present in a waste dump of a Cu mine with a high level of natural radionuclides. The results obtained using a NN were compared with those estimated by Co-Kriging. Both models reproduced field measurements equivalently as a function of spatial coordinates. Similarly, the deviations from the reference concentration values obtained in the output layer of the NN were smaller than the deviations obtained from the multiple regression analysis (MRA), as indicated by the results of the root mean square error. Finally, the method validation showed that the estimation of radiological parameters based on their spatial coordinates faithfully reproduced the affected area. The estimation of the activity concentrations was less accurate for both the NN and MRA; however, both methods gave statistically comparable results for activity concentrations obtained by gamma spectrometry (Student\'s t-test and Fisher\'s F-test).
摘要：
受天然放射性核素污染的土壤的放射性特征可以对所调查的区域进行分类，优化实验室测量，并就潜在的场地修复做出明智的决策。神经网络（NN）正在成为执行这些任务的新候选者，作为传统地质统计学工具（如协同克里格法）的替代品。这项研究证明了神经网络的实施，用于估计放射学值，如环境剂量当量(H*(10))，天然放射性核素含量高的铜矿废物堆中存在的天然放射性核素的表面活性和活性浓度。将使用NN获得的结果与联合克里格法估计的结果进行比较。两种模型都等效地将现场测量结果再现为空间坐标的函数。同样，在NN的输出层中获得的与参考浓度值的偏差小于从多元回归分析（MRA）获得的偏差，如均方根误差的结果所示。最后,方法验证表明，基于其空间坐标的放射学参数估计忠实地再现了受影响的区域。对于NN和MRA，对活性浓度的估计都不太准确；但是，两种方法对通过γ能谱(Student'st检验和Fisher'sF检验)获得的活性浓度给出了统计学上可比较的结果.
