基于通道注意力和代价敏感改进图像分割的矿区环境遥感监测

Remote Sensing Monitoring of Mining Area Environment Based on Image Segmentation Improved by Channel Attention and Cost Sensitivity

为了实现对矿区遥感图像分割,从而判别矿区时空环境变化,达到矿区环境遥感监测的目的,提出了改进语义分割的矿区环境监测模型。首先将U-Net作为语义分割模型的基础网络,其次为了增强模型对小目标的学习能力,使用了代价敏感权重向量方法,改进了训练时的目标函数,并且利用通道注意力模型挖掘卷积特征图的通道依赖性,最后通过条件随机场将U-Net的分割结果精细化。相对于传统算法,改进后的算法极大提升了矿区环境小目标的边缘分割精度,便于从遥感图像上更加准确地监测矿区环境变化,如房屋移位、水体变化等,为矿区安全建设和开采计划调整提供决策参考依据。

For realizing the remote sensing image segmentation of mining area,so as to distinguish the change of mining area space-time environment and to achieve the purpose of remote sensing monitoring of mining area environment,a mining area environment monitoring model based on improved semantic segmentation was proposed.Firstly,U-Net was used as the basic network of the semantic segmentation model.Secondly,in order to enhance the learning ability of the model for small targets,the cost sensitive weight vector method was used to improve the objective function during training,and the channel dependence of convolution feature graph was developed by channel attention model.Finally,the segmentation result of U-Net was refined by conditional random field.Compared with the traditional algorithm,the improved algorithm greatly improves the edge segmentation accuracy of small targets in the mining area,which is convenient to monitor the environmental changes of mining area more accurately from remote sensing images,such as house displacement,water body change.The study can provide a decision-making reference for safety construction and mining plan adjustment of mining area.