基于融合注意力机制的小样本遥感场景分类方法

Small-sample remote sensing scene classification method based on fused attention mechanism

针对遥感场景图像样本获取困难,数据量受限以及遥感图像目标对象和背景高度混杂的问题,提出一种基于融合注意力机制的小样本遥感场景分类方法。该方法采用RepVGG作为基准模型,并利用ECANet网络的ECA通道注意力机制改进RepVGG网络的RepVGGBlock模块,使得网络有效过滤无用信息并聚焦于关键场景区域,从而增强模型的特征判别能力,并确保在不增加模型参数的情况下提高分类准确率;同时通过随机数据增强方法在线增强训练数据,在不占用额外内存的情况下增加模型训练数据量,使得训练数据更多样化,提高模型的泛化能力。在UC Merced LandUse数据集上分类平均准确率为94.52%,相较于ResNet50、RepVGG-B1-SE网络,准确率分别提高4.52%和2.93%。实验结果表明,该方法能有效聚焦关键场景区域并提升小样本遥感场景分类的准确率,对实现遥感影像快速分类具有一定的参考意义。

The problems of difficult sample acquisition may lead to the limited data volume for remote sensing scene.The highly mixed target objects and backgrounds of remote sensing images also bring difficult problems of remote sensing scene classification.Aiming at the above problems,a small sample remote sensing scene classification method based on a fused attention mechanism is proposed.In the proposed method,the ECA channel attention mechanism of ECANet network is employed to improve the RepVGGBlock model of the RepVGG network.Within the integrating fused attention mechanism,the proposed network can effectively filter useless information and focus on key scene regions to enhance the feature discrimination ability of the model and ensure the classification accuracy without increasing the model parameters.Meanwhile,the training data is enhanced online by random data augmentation method to increase the amount of model training data without taking up extra memory,which makes the training data more diverse and improves the generalization ability of the model.The average accuracy of classification on the UC Merced LandUse dataset is 94.52%,which is 4.52%and 2.93%points higher than that of ResNet50 and RepVGG-B1-SE networks,respectively.The experimental results show that the method can effectively focus on key scene regions and improve the accuracy of small-sample remote sensing scene classification,which will be benefit for realizing fast classification of remote sensing images.