基于通道域注意力卷积神经网络的遥感影像高尔夫球场提取

Golf Course Land Extraction Based on Channel Attention Convolution Neural Network

针对高尔夫球场目标大、场景复杂等引起的难以完整准确提取问题,构建了EfficientNetB3+UNet网络,引入通道域注意力模块,设计了大尺寸样本训练策略,并对模型进行了对比实验分析和应用。实验结果表明,所提方法的mIoU精度为0.948 7,明显高于非通道域注意力模型(mIoU为0.884 8)和小尺寸模型(mIoU为0.601 4),有效提升了高尔夫球场提取的完整性和准确性,显著降低了自然植被、水域等复杂内部场景和球场边缘混合场景等导致的漏提和误提现象。同时,在北京、上海、广州和深圳四地的高尔夫球场提取应用中,模型召回率均优于90%,具有良好的应用价值。

In view of the difficulty of complete and accurate extraction caused by large golf course target and complex scene,the EfficentNetB3-UNet network is constructed,the channel domain attention module is introduced,the training strategy of large-size sample is designed,and the model is compared and applied.The results show that the mIoU accuracy of this method is 0.948 7,which is significantly higher than that of the non-channel domain attention model(mIoU is 0.8848)and the small input size model(mIoU is 0.601 4).This method effectively improves the integrity and accuracy of golf course extraction,and significantly reduces the leakage and mispronouncing caused by complex internal scenes such as natural vegetation,water and mixed scenes at the edge of the course.In the golf courses comparison among Beijing,Shanghai,Guangzhou and Shenzhen,Beijing has the largest number of golf courses and Shenzhen has a significantly higher density of courses than the other three cities.