 Our proposed method solves the problem of overfitting and lack of depth in CNNs for remote-sensing scene classification. It uses a dense connection architecture to create a deep network with fewer parameters while maintaining high accuracy. Additionally, it employs adaptive average 3D pooling to reduce the number of channels and fix the size of feature maps. Moreover, it replaces the fully connected layer with a convolutional layer, allowing the output features to be directly classified without flattening operations. Finally, our algorithm outperforms existing methods on four datasets, demonstrating its effectiveness. This article was offered by Jianming Zhang, Chaokuan Lu, Sudang Li, and others.