Remote Sensing Image Scene Classification Based on ResNet50 and Channel Attention
Convolutional neural networks(CNNs)are widely used for image classification.Howev-er,directly applying traditional CNNs to remote sensing image scene classification has certain limita-tions.Remote sensing images exhibit large intra-class differences and high intra-class similarities,making it challenging for networks to accurately extract image features,thus posing significant diffi-culties in remote sensing image classification tasks.The channel attention mechanism,with its abili-ty to focus on important features while ignoring less relevant ones,can enhance the recognition capa-bility of CNNs for image features.Therefore,a network model combining ResNet50 and channel at-tention,called ResNet50+Attention,is proposed.The UCMD data set is used with the initialization of network parameters for remote sensing scene image classification tasks.Classic network models in-cluding AlexNet,DenseNet,VGG16,and GoogLeNet are compared.ResNet50+Attention outper-forms other models significantly in terms of overall accuracy,precision,recall,and specificity clas-sification metrics.Disintegration experiments with the base ResNet50 model are conducted,inclu-ding accuracy curves,confusion matrices,and individual class classification metrics.The results show that ResNet50+Attention achieves overall accuracy,precision,recall,and specificity of 91.7%,92.1%,91.8%,and 99.6%,respectively,showing improvements of 4%,3.8%,4%,and 0.2%compared to ResNet50,thus confirming the effectiveness of this network model.
国家科技期刊平台
NSTL
万方数据
维普
