The excellent performance of convolutional neural network (CNN) and its variants in image classification makes it a potential perfect candidate for dealing with multi-geoinformation involving abundant spatial information. In this paper, we tested, for data-driven mineral prospectivity mapping, the efficacy of using unsupervised convolutional auto-encoder network (CAE) to support CNN modeling for synthesis of multi-geoinformation. First, two simple unsupervised CAE networks were constructed to distinguish patches of tif image (i.e., nine predictive evidence maps forming a tif-format image) with nine channels that have high reconstructed errors, which represent prospective areas (i.e., mineralized). Then, the patches of tif image with the lowest reconstructed errors were regarded as background (or non-prospective areas). We varied the CAE network architecture and training epochs and combinations of evidence maps for trials to obtain reliable results. Then, the AUC, or area under the receiver operating characteristic curve, was used to demonstrate empirically that high reconstructed errors are representative of spatial signatures of prospective areas. The proposed coherent spatial signatures, namely patches of a tif image with the highest reconstructed errors and the lowest reconstructed errors representing prospective and non-prospective areas, respectively, were used in the subsequent CNN modeling. The results of CNN modeling using training data derived from CAE exhibited strong spatial correlation with known Au deposits in the study area. The training loss and accuracy of the CNN modeling together with resulting favorability map that were comparable with results from previous study proved the plausibility of the proposed methodology, and therefore, the practice of extracting coherent spatial signatures of prospective and non-prospective areas in unsupervised manner using CAE network and then using these coherent spatial signatures in supervised learning with CNN is a new potential approach for mineral prospectivity mapping.

卷积神经网络（CNN）及其变体在图像分类中的出色表现使其成为处理涉及大量空间信息的多地理信息的潜在理想人选。在本文中，我们针对数据驱动的矿物远景测绘测试了使用无监督卷积自动编码器网络（CAE）支持CNN建模以合成多种地理信息的功效。首先，构建了两个简单的无监督CAE网络，以区分九个具有较高重构误差的通道的tif图像斑块（即形成tif格式图像的九个预测证据图），这些通道代表预期区域（即矿化区域）。然后，将具有最小重构误差的tif图像块视为背景（或非预期区域）。我们改变了CAE网络架构和训练时期，并结合了证据图进行试验以获得可靠的结果。然后，使用AUC或接收器工作特性曲线下的区域，从经验上证明高重建误差代表了预期区域的空间特征。所提出的相干空间特征，即分别代表预期区域和非预期区域的tif图像分别具有最高重构误差和最低重构误差的补丁，被用于后续的CNN建模中。使用源自CAE的训练数据对CNN建模的结果显示出与研究区域中已知的金矿床有很强的空间相关性。