Mine Remote-Sensing Datasets

While providing resources for socio-economic development, mine development has also given rise to problems such as disorderly and illegal mining, and has caused serious damage and pollution to the surface environment, among a series of sustainable development problems. Therefore, remote sensing interpretation of mine development (including mine scenes and mine targets) is automatically realized with the help of machine learning and deep learning techniques, which can provide important support for the assessment of sustainable development of mining areas at the regional scale. However, the current lack of high-precision mine scenes and target datasets at a wide-area scale restricts the development of intelligent interpretation of remote sensing for mine development. In this paper, a mine multi-scene classification dataset, as well as a mine target detection dataset, are constructed in Hubei and Jiangxi provinces in the middle reaches of the Yangtze River. The dataset is characterized by a wide coverage area, multiple mine types, high image resolution, and multi-scale. In particular, for the mine multi-scene classification dataset, a feature fusion mine scene classification method based on a stereo attention mechanism is proposed, in which the salient features of the mine scene in three directions, horizontal, vertical, and channel, are modeled by the constructed stereo attention mechanism, and a feature activation module is used to avoid generating feature disappearance; For the mine target detection dataset, a mine target detection method based on a multi-scale dual-attention mechanism is proposed, which avoids feature bias through a multi-scale feature fusion module, while using a dual-attention mechanism to highlight salient features of the mine. The experimental results of the two models show that the mine multi-scene and target detection dataset constructed in the paper has good quality and can provide a benchmark dataset for the intelligent interpretation research of remote sensing for mine development in the Yangtze River Basin and even and global scale.

矿产开发在为社会经济发展提供资源支撑的同时，也催生了无序开采、非法采矿等诸多问题，并对地表环境造成了严重破坏与污染，衍生出一系列可持续发展层面的难题。因此，借助机器学习与深度学习技术实现矿产开发遥感解译（涵盖矿山场景与矿山目标），可为区域尺度矿区可持续发展评估提供重要支撑。然而当前缺乏广域尺度下的高精度矿山场景与目标数据集，制约了矿产开发遥感智能解译技术的发展。为此，本文以长江中游地区的湖北省与江西省为研究区域，构建了矿山多场景分类数据集与矿山目标检测数据集。该数据集具备覆盖范围广、矿山类型多样、图像分辨率高且具备多尺度特性等优势。针对矿山多场景分类数据集，本文提出了一种基于立体注意力机制（Stereo Attention Mechanism）的特征融合矿山场景分类方法：通过构建的立体注意力机制对矿山场景在水平、垂直与通道三个维度的显著特征进行建模，并引入特征激活模块以避免特征消失问题；针对矿山目标检测数据集，本文提出了一种基于多尺度双注意力机制（Multi-scale Dual-attention Mechanism）的矿山目标检测方法：通过多尺度特征融合模块规避特征偏移问题，同时利用双注意力机制强化矿山目标的显著特征。两种模型的实验结果表明，本文所构建的矿山多场景与目标检测数据集具备优良的质量，可为长江流域乃至全球尺度下的矿产开发遥感智能解译研究提供基准数据集。