结合场景分类数据的高分遥感图像语义分割方法

SEMANTIC SEGMENTATION OF HIGH RESOLUTION REMOTE SENSING IMAGES BY INTEGRATING SCENE CLASSIFICATION DATA

深度语义分割模型作为解决图像像素级分类的重要方法,在遥感图像分类中的应用受到遥感像素级标记数据不足的制约,在有限数据的情况下训练后的网络难以有效提取遥感图像特征。为此,将具有图像级标记的遥感场景分类数据应用到语义分割模型训练中。利用遥感场景分类数据训练卷积神经网络模型,并以其为基础构建语义分割网络的特征提取部分,从而提高语义分割模型提取遥感图像特征的能力。在训练卷积神经网络的过程中,对训练数据基于后验概率进行类别映射与平衡,使其更贴近目标任务的遥感图像。实验结合场景分类数据集UC Merced Landuse训练语义分割模型,在高分辨率遥感数据集Potsdam上获得了89.50%的总体分类准确率,证明该方法提高了语义分割模型在遥感数据上的像素级分类效果。

The deep semantic segmentation model is an important approach to solve the pixel-level classification problem for images.However,its application in remote sensing image classification is limited by the insufficiency of remote sensing images with pixel-level labels.The network can t well extract features for remote sensing images in the cases where there are limited training data.Therefore,remote sensing scene classification data with image-level labels are utilized in the semantic segmentation model training.These data were used to train a convolutional neural network,and the feature extraction part of a semantic segmentation model was constructed based on the trained network so that the semantic segmentation model could better extract features for remote sensing images.While training the convolutional neural network,the classes of training data were mapped and balanced based on posterior probabilities to make them closer to the target remote sensing images.In the experiment,semantic segmentation models were trained with the integration of a scene classification dataset UC Merced Landuse.The model gains an overall accuracy of 89.50%on the high resolution remote sensing dataset Potsdam,validating that our method can improve the classification effect of semantic segmentation model on remote sensing data at pixel level.