基于多通道自注意力网络的遥感图像场景分类

Remote sensing image scene classification based on multi-channel self-attention network

高分辨率遥感图像场景分类广泛应用于土地监测、环境保护及城市规划等诸多领域.现有场景分类方法不能很好地结合局部纹理信息和全局语义信息,同时各通道特征之间的关系没有得到有效挖掘.因此,本文提出了一种基于多通道自注意力网络的遥感图像场景分类模型.通过卷积网络提取遥感图像的多尺度特征;随后采用特征融合单元建立多尺度特征间的局部-全局关系,基于多头自注意力机制的Inter-Channel Transformer在通道维度对融合后的特征建模,并推导特征在通道间的关系,进一步扩大全局感受野,以捕捉其语义结构信息,有效提高了网络的分类精度.在数据集AISC和SIRI-WHU上,本文所提算法的整体分类准确率(OA)分别为95.70%和94.00%,超过了当前最新的研究算法,证明了所提模型在高分辨率遥感图像场景分类任务中的有效性.

High resolution remote sensing image(HRRSI)scene classification is widely used in many fields such as land monitoring,environment protection,urban planning and so on.The existing scene classification methods cannot fuse the local-texture and global-semantic information well,and the relationship between the features of each channel has not been effectively explored.Therefore,this paper proposed a new method based on multi-channel self-attention network for HRRSI scene classification.Firstly,the multi-resolution features are extracted by Convolutional Neural Network(CNN);then,a feature fusion unit is used to establish the local-global relationship between multi-scale features.In addition,Inter-Channel Transformer,which is based on multi-head self-attention mechanism,models the merged representations in the channel dimension,and reasons the relationship between the features of each channel,further expands the global receptive field to capture its semantic structure information.Finally,the proposed method improves the classification accuracy.This paper also designs series of experiments on AISC and SIRI-WHU datasets to demonstrate the validity of the proposed algorithm for HRRSI scene classification task.The OA(Overall Accuracy)performance are 95.70%and 94.00%on AISC and SIRI-WHU respectively.It has surpassed the state-of-the-art algorithms.