基于多尺度特征和元学习的智能预测找矿靶区实验研究

当前智能找矿靶区预测方法大多依赖于人工采样和专家的知识经验，然而，对于现实世界中区域小、数量少的矿区区域，这些方法将面临巨大的挑战。为了迎接这个挑战，提出一种新颖的深度智能找矿靶区预测框架——多尺度特征交互框架。具体地，首先定义两个网络，即多尺度特征映射网络和多尺度特征分类网络；在此基础上，通过膨胀卷积捕获多尺度特征映射网络中不同地球化学元素的特征，并且利用多尺度分类网络处理这些特征；其次，使用元网络为多尺度分类网络生成卷积权重；最后使用自蒸馏挖掘多尺度分类网络中的隐知识用于预测。整个模型采用端到端的训练方式，大量的实验结果表明，多尺度特征交互框架与当前最先进的方法比较具有显著的竞争力。

Experimental research on intelligent prospecting prediction based on multi-scale feature and meta-learning

At present, most prediction methods of prospecting prediction rely on manual sampling and experts'' knowledge and experience. However, these methods face great challenges in the real world with small areas and less numbers of mines. In order to meet this challenge, this paper proposed a novel depth prospecting prediction framework: multi-scale feature interaction framework. Firstly, this paper defined two networks, the multi-scale feature mapping net and the multi-scale feature classification net. On this basis, it captured the features of different geochemical elements in the multi-scale feature mapping net by dilated convolution, and used the multi-scale feature classification net to process these features. Secondly, it used meta network to generate convolutional weights for multi-scale classification networks. Finally, it used self-distillation to exploit implicit knowledge in multi-scale classification networks for prediction. The whole model adopted end-to-end training mode. A large number of experimental results show that multi-scale feature interaction framework is significantly competitive with the most advanced methods.