This paper proposes an approach to remote sensing scene classification based on the DenseNet-161 convolutional neural network and transfer learning strategies. Experiments conducted on international benchmark datasets such as UC Merced, AID, and NWPU-RESISC45 demonstrated that the DenseNet-161 model achieved high accuracy rates: 99.29% on UC Merced, 96.15% on AID, and 94.03% on NWPU-RESISC45. The model’s performance was compared with popular architectures such as ResNet, Inception, and VGG, with DenseNet-161 showing clear advantages, especially on complex and resource-constrained datasets. The study also provides a practical analysis of how transfer learning can enhance the generalization and efficiency of deep learning models. The obtained results confirm that DenseNet-based transfer learning solutions are effective for automatic and