基于语义分割深度学习的找矿靶区预测研究——以广东省阳江-茂名地区为例

找矿靶区预测需要综合考虑地质背景、地球化学数据、地球物理勘探数据、遥感数据等因素。随着人工智能时代的到来，靶区预测可以最大限度地利用计算机运算性能，通过特定的规则集成所有地学数据对各类矿种的找矿靶区进行预测，尽可能规避由于数据种类多、数据量大、方法复杂、主观性强造成的预测结果可靠性差等问题。本文以广东省阳江-茂名地区为例，融合地球化学、地层岩性、地质构造、地形地貌等数据，基于PSPNet、SegNet、UNet三种语义分割深度学习模型进行预测，结果表明PSPNet模型在预测精度方面优于SegNet及UNet模型，并预测出了55处铁矿、金矿、铜矿、高岭土矿找矿靶区，其中79.7%的已查明矿点位于预测靶区内，表明该方法在找矿靶区预测中具有较高的可行性，可以用于找矿勘查并圈定靶区。

Prediction of prospecting targets based on semantic segmentation with deep learning:A case study of the Yangjiang-Maoming area in Guangdong Province

The prediction of prospecting target areas needs to take into account the geological background, geochemical data, geophysical exploration data, remote sensing data and other factors. With the arrival of the era of artificial intelligence, target prediction can maximize the utilization of computer computing performance, integrate all geological data through specific rules to predict ore prospecting targets, and try to avoid problems such as poor reliability of prediction results due to multiple types of data, large amount of data, complex methods, and strong subjectivity in target area prediction process. This work took Yangjiang-Maoming area in Guangdong Province as an example, integrated geochemistry, lithology, geological structures, landform and other data, and predicted targets based on three semantic segmentation models with deep learning: PSPNet, SegNet, and UNet. The results indicate that the PSPNet model outperforms SegNet and UNet models in terms of prediction accuracy, and has predicted 55 target areas for iron ores, gold ores, copper ores, and kaolinite ores. Among them, 79.7% of the identified ore spots are located within the predicted target areas, indicating that this method has high feasibility and can be used for ore exploration and target delineation.