基于深度学习的遥感图像语义分割预测增强技术

Remote sensing image semantic segmentation prediction augmentation based on deep learning

深度学习是实现遥感图像自动解译的重要途径,但遥感图像尺度往往较大,需要剪裁成小尺度图像进行训练预测,而不同的预测策略将影响最终的预测效果.本文论述了遥感图像语义分割的4种预测策略,并通过在国际摄影测量与遥感学会(International Society for Photogrammetry and Remote Sensing,ISPRS)波茨坦数据集上训练全卷积网络(fully convolutional network,FCN)、U-Net、SegNet和DenseASPP模型,对这些预测策略进行了评估.研究结果显示,使用预测增强技术相比不使用预测增强大约能提高2%~3%的分割精度,但随之而来的是需要消耗更多的计算资源.

Deep learning is an important way to realize automatic interpretation of remote sensing images.However,the scale of remote sensing images is often large,such large-scale images need to be cropped into small-scale images for training and prediction,and different prediction strategies will affect the final prediction effect.In this paper,four prediction strategies for remote sensing image semantic segmentation are discussed,and fully convolutional network(FCN),U-Net,SegNet and DenseASPP models are trained on the International Society for Photogrammetry and Remote Sensing(ISPRS)Potsdam dataset to evaluate these prediction strategies.The research result shows that the use of predictive enhancement technology can improve the segmentation accuracy by about 2%-3%compared with not using predictive enhancement,but it will consume more computing resources.