融合全局上下文的近岸养殖池塘提取算法

Inshore aquaculture pond extraction algorithm based on global context fusion

针对现有方法对于养殖池塘和干扰地物的区分效果不足,在多源高分辨率遥感影像上的普适性有待验证等问题,提出一种融合全局上下文信息的PG-Unet养殖池塘提取模型.该模型在U-Net的基础上,通过增加金字塔特征提取单元来捕捉丰富的全局上下文信息,增加全局引导流来改善不同级别特征图的质量,提升模型在多干扰地物环境定位目标的能力.在GF-2 PMS和BJ-2 PMS数据集上的实验结果表明,PG-Unet模型精度最优,其IoU和F_(1)分数分别达到92.30%、96.00%和92.07%、95.87%,优于U-Net、DensenetUnet和U2Net等方法,具有更强的抗干扰能力和普适性,能更好地区分养殖池塘和干扰地物;同时,PG-Unet模型在诏安湾养殖区域应用也取得了较高的提取精度,能够实现大范围养殖池塘空间分布信息自动精准提取.

Aiming at the problems that the existing methods are not effective in distinguishing aquaculture ponds and interfering objects,and the universality of multi-source high-resolution remote sensing images needs to be verified,a PG-Unet aquaculture pond extraction model integrating global context information is proposed.On the basis of U-Net,the model captures rich global context information by adding pyramid feature extraction unit,and increases global guiding flow to improve the quality of feature maps at different levels,so as to improve the ability of the model to locate targets in multi-interference environment.The experimental results on GF-2 PMS and BJ-2 PMS datasets show that the PG-Unet model has the best accuracy,and its IoU and F_(1)scores reach 92.30%,96.00%and 92.07%,95.87%,respectively,which are better than U-Net,DensenetUnet and U2Net.It has stronger anti-interference ability and universality,and can better distinguish aquaculture ponds and disturbed objects.At the same time,the application of PG-Unet model in Zhao’an Bay aquaculture area has also achieved high extraction accuracy,which can realize the automatic and accurate extraction of spatial distribution information of large-scale aquaculture ponds.