一种基于TransUNet的SAR影像水体提取及洲滩面积变化监测应用

Water extraction from SAR images based on TransUNet and its application in sandbar area change monitoring

针对面向合成孔径雷达(SAR)影像的高精度水体提取工作仍然存在较大的挑战,该文基于TransUnet深度学习模型和COSMO-SkyMed卫星的高分辨率(3 m)数据,深度挖掘Transformer模型的全局上下文捕捉能力和U-Net模型多尺度特征提取优势,建立了小样本数据集下的SAR影像水体提取模型。实验结果表明该文算法在小面积水体提取、山体阴影误判抑制方面优势明显,其水体提取结果的准确率、召回率、总体准确率、F1分数、交并比分别为89.54%、91.24%、98.01%、90.26%和82.28%,相较于U-Net模型、FCN-VGG16模型和HRNet模型,精度具有较大提升;同时采用该文的水体提取模型反演了洞庭湖区2019年7月—2020年6月的7景SAR影像的洲滩空间分布变化情况,阐明了其随季节性水位变化的年周期变化规律。

Aiming at the challenges for high-precision water body extraction from synthetic aperture radar(SAR)images,the TransUnet deep learning model and high-resolution(3 m)data from COSMO-SkyMed satellites were used to delve into the global contextual capturing capabilities of the Transformer model and the multi-scale feature extraction advantages of the U-Net model in this paper.The water body extraction model for small sample datasets of SAR images was established.Experimental results revealed pronounced advantages of the proposed algorithm in extracting small-area water bodies and mitigating misclassifications in mountainous shadow regions.The accuracy,recall,overall accuracy,F1 score,and intersection over union(IoU)of the water body extraction results were reported as 89.54%,91.24%,98.01%,90.26%,and 82.28%,respectively.These results showed a noticeable improvement compared to the Unet,FCN-VGG16,and HRNet models.Concurrently,the proposed water body extraction model was utilized to infer the spatial distribution changes of sandbars in the Dongting Lake region across seven SAR images captured from July 2019 to June 2020.This elucidated the annual cyclical variation in response to seasonal water level changes.