Remote sensing image scene recognition based on densenet-169

Scene classification has become an effective technique for classifying high spatial resolution remote sensing images. However, in the traditional deep learning convolutional neural network, as the remote sensing image passes through the layers in the network, some of the features will be gradually lost, resulting in a significant decrease in the classification accuracy and precision of scene recognition, and there is the problem of underutilization of features. In addition, remote sensing images themselves have high complexity. To overcome these challenges, we adopt the DenseNet network. Specifically, we first train remote sensing images from the UCMerced dataset as network inputs. Then, we introduced the DenseNet-169 model based on migration learning. Compared with DenseNet-121, DenseNet-169 has more neural network layers, and this difference is mainly manifested in the number of convolutional layers in the dense blocks.DenseNet-169 has more convolutional layers, which increases the complexity and the number of parameters of the model, bringing the following advantages: stronger expressive power, which enables the extraction of more complex feature patterns; faster training time, thanks to the densely-connected nature, which efficiently utilizes the gradient flow; and better generalization ability, especially on large-scale complex datasets. In our experiments, the introduced DenseNet-169 shows excellent performance compared to other state-of-the-art deep convolutional networks on the UCMerced dataset, with an accuracy of 95.14%, a precision of 95.31%, a Kappa coefficient of 94.90%, and an F1-score of 95.11%. The experimental results show that the method can make full use of the features of remote sensing images and show good visual effect, providing a good method for scene recognition.