机器学习在水系沉积物地球化学采样下伏基岩填图中的应用

Mapping of underlying bedrock based on the stream sediment geochemical samples by applying the machine learning algorithms

为了探索高效且能大区域应用的基岩填图方式,本文整合水系沉积物地球化学数据和局部空间Moran’s I指数、高程、坡度、坡向变率、高磁、断层、矿点、水系等多元地学数据作为附加特征,采用浅层机器学习决策树及其Bagging和Boosting集成算法,以及深度学习的图卷积网络,分别训练了不同的基岩判别模型。结果表明,相比浅层机器学习的决策树及其集成算法,深度学习的图卷积神经网络基岩判别模型仅使用20%带标签数据就获得了最高的78.31%判别精度。应用基于图卷积网络的基岩类型判别模型对察汗乌苏河地区第四系覆盖物下伏基岩填图,预测结果与其周边基岩类型协调一致,该模型可用来探究更全面的区域基岩分布情况。

In order to explore effective bedrock mapping methods which can be applied in a big area, by integrating data of the stream sediment geochemical survey with data of the local spatial autocorrelation Moran’s I indicator, terrain characteristics including elevation, slope and slope of aspect, high resolution magnetic survey, fault, ore occurrence, and stream as the additional characteristics, and by using the decision tree(DT) and its Bagging ensemble and Boosting ensemble algorithms, belonging shallow machine learning, and the graphic convolutional network(GCN), belonging deep learning, we have respectively trained five different models for recognizing the bedrock types underlying the geochemical sampling locations in this paper. The experimental results show that the highest precision of 78.31% for recognizing bedrock types has been obtained by using the GCN algorithm for calculating data of only 20% labeled training samples comparing with that obtained by using the DT and its ensemble algorithms for calculating all data. Using the GCN bedrock classification deep learning model, we have carried out an mapping of the bedrock types underlying the Quaternary coverages in the Chahanwusu River area, Qinghai Province. The predicted bedrock types are consistent with those of verified neighboring bedrocks. Therefore, the GCN bedrock classification deep learning model can be used to fully explore the spatial distribution of bedrocks in the big region.