改进Oriented R-CNN的遥感尾矿库检测

An Improved Oriented R-CNN for Tailings Ponds Detection on Remote Sensing Images

尾矿库中含有大量的尾砂,是一个具有高势能的人造泥石流危险源,一旦发生溃坝危险,就会带来严重的人员损失和环境灾难。掌握尾矿库的数量和空间分布情况,对尾矿库事故的预防具有重要的意义。传统的尾矿库调查依赖人工目视解译和地面验证,难以实现大范围、高频次的监测。文章以多源高分辨率卫星影像为数据源,采用人工遍历解译的方式标注样本,并结合多种数据增广方法构建了深度学习的尾矿库检测数据集。在此基础上,通过嵌入轻量化的注意力机制模块,同时设置自适应的锚框,优化Oriented R-CNN模型。实验结果表明:改进后的模型在尾矿库检测数据集上的性能显著提升,全类平均精度和召回率分别能够达到84.14%和90.32%,同时模型具有较强的可靠性和泛化性。文章提出的方法有利于推动尾矿库自动化、智能化的应急监管。

The accidents of tailings pond may lead to casualties and environmental pollution.It is of great significance to monitor tailings pond timely and accurately for the prevention and management of dams’accidents.Traditional remote sensing methods,relying on manual visual interpretation and ground verification,are inefficient and unsuitable for large scale extraction.Using multi-source high-resolution satellite images as data source,a high quality dataset for tailings pond detection is constructed by manual interpretation,combining a variety of data augmentation methods.On this basis,the Oriented R-CNN model is optimized by embedding a lightweight attention mechanism and feature pyramid network to residual structure,which is effectively to fuse deep multi-scale features and recalibrate the input contributions.Meanwhile,a series of original anchors are generated adaptively according to the self-made dataset.The experimental results show that the mean average precision and recall of the proposed algorithm for tailings pond detection reache 84.14%and 90.32%,respectively,which indicate a significant improvement compared with the original Oriented R-CNN.The results also prove the feasibility and generalization of the improved model.This study is important for large-scale,high-precision,and intelligent monitoring of tailings ponds,which can serve as a reference for tailings pond management.