摘要
当前大比例尺(1∶5万及以上)找矿预测场景中普遍存在样本构建过程与成矿系统理论脱节、样本集非均衡分布和模型训练效率低等问题,在实际应用中容易导致异常识别精度下降、隐伏矿体漏检率升高和预测结果可解释性减弱,制约了快速找矿发现和勘查突破。本研究以腾冲-小龙河锡矿集区为示范,建立了成矿系统理论引导的样本构建机制;在此基础上,创新性设计深-浅层混合架构模型,通过深层残差卷积网络捕捉复杂非线性特征,结合浅层单分类支持向量机异常检测器实现了多源地学信息高效融合与矿化异常快速识别,有效缓解了样本极度不平衡的问题。预测结果表明,高概率找矿靶区不仅与86.7%的已知矿床(点)具有显著空间吻合性,也与燕山期岩体围岩接触带及云英岩脉发育区呈现显著耦合关系。据此成功圈定了11处找矿远景区,并在找矿中心实施的验证钻孔中成功揭露了锡钨矿化。上述成果从地质规律与工程实证两方面共同印证了残差卷积-异常检测混合架构不仅预测可靠、实用价值显著,同时在保持与深度自编码网络相当的非线性表征能力基础上具备更高的训练效率,显示出良好的应用潜力。
In current large-scale(1∶50000 or finer)mineral prospectivity modeling scenarios,several critical challenges persist,including disconnection from metallogenic theories,non-uniform sample distribution,and low model training efficiency.Theseissues often result in reduced anomaly detection accuracy,an increased risk of missing concealed ore bodies,and diminishedinterpretability of prediction results in practical applications,thereby hindering rapid mineral discovery and explorationbreakthroughs.This research focused on the Tengchong-Xiaolonghe tin polymetallic ore district,and a metallogenic system theory-guided sample construction mechanism to harmonize domain knowledge with data-driven approaches was established.On this basis,a novel deep-shallow hybrid architecture was developed.The model leverages deep residual convolutional networks to capturecomplex nonlinear features,and integrates shallow one-class support vector machines to enable efficient fusion of multi-sourcegeoscientific data and rapid identification of mineralization anomalies.This approach effectively alleviates challenges related toimbalanced sample distribution and low training efficiency under large-scale scenarios.The prediction results indicate that the high-probability prospecting targets exhibit a strong spatial correspondence with 86.7%of the known deposits(occurrences)and show apronounced coupling relationship with the contact zones between Yanshanian granitoids and wall rocks,as well as with thedistribution of greisen veins.Accordingly,11 prospective mineralization zones were delineated,and inspection holes conductedwithin these zones successfully intercepted Sn-W mineralization.These results,corroborated by both geological evidence andengineering verification,confirm that the residual convolution-anomaly detection hybrid architecture provides reliable and practicallymeaningful predictions.Moreover,it achieves higher training efficiency while maintaining nonlinear representation capabilitycomparable to that of deep autoencoder networks,demonstrating strong potential for practical applications.
作者
周放
张玙
周清
马龙
梁虹
韩志婷
ZHOU Fang;ZHANG Yu;ZHOU Qing;MA Long;LIANG Hong;HAN Zhiting(Chengdu Center,China Geological Survey(Geosciences Innovation Center of Southwest China),Chengdu 610218,China)
出处
《沉积与特提斯地质》
北大核心
2025年第4期737-750,共14页
Sedimentary Geology and Tethyan Geology
基金
中国地质调查局地质调查项目(DD20240070,DD20242602,DD20240069,DD20242494)
深地国家科技重大专项(2024ZD100320703)。
关键词
智能找矿预测
腾冲-小龙河锡矿
多源地学信息融合
异常检测
深-浅层混合神经网络
intelligent mineralization prediction
Tengchong-Xiaolonghe tin deposit
multi-source geoscientific data fusion
anomaly detection
deep-shallow hybrid neural network
