西南多源遥感影像水资源监测方法研究

Research on Water Resource Monitoring Method Using Multi-source Remote Sensing Images in Southwest China

为解决西南地区水资源监测的问题,结合遥感影像的近红外波段,对多源遥感影像进行全局线性拉伸,全局自适应gamma变换等预处理。提出了一种基于SegFormer水体提取方法,引入了注意力机制,提取不同分辨率下的水体语义信息,实现水体的高精度像素级分割,从而快速自动的监测西南地区水资源分布情况。以四川绵阳作为实验区,利用国家基础地理信息数据构建多源遥感影像样本数据集进行模型训练,对实验区的国产遥感卫星影像进行水体提取,实验结果显示:最高正确率为95.6%,召回率为97%,IOU为0.929。结果证明该方法的水体提取相较于其他方法具有更好地精度和鲁棒性。

In order to solve the problem of water resource monitoring in southwest areas,we proposed a water body extraction method based on SegeFormer,combining the near infrared band of remote sensing image with the pre-processing of linear stretching and gamma transformation of multi-source remote sensing images.We introduced the attention mechanism to extract the semantic information of water at different resolutions,which could achieve high-precision pixel-level segmentation of water,thus quickly and automatically monitoring the water resource distribution in the southwestern areas.Taking Mianyang,Sichuan as the experimental area,we used national basic geographical information data to build multi-source remote sensing image sample dataset for training,and carried out water extraction of the domestic remote sensing satellite images in the experimental area.The results show that the highest accuracy rate is 95.6%,the recall rate is 97%,and the IoU is 0.929,which illustrating this method has better accuracy and robustness than other methods in water extraction.