Abstract: Addressing the challenge posed by the uneven distribution of various features and the low classification accuracy of urban remote sensing images, we propose a novel method for remote sensing image classification that integrates parallel attention and weight balance algorithm. Leveraging the semantic segmentation network architecture of DeepLabV3+ and ResNet50, our method combines channel attention and spatial attention algorithms in parallel to improve the network's feature extraction capability. Additionally, to address the issue of imbalanced feature category proportions in remote sensing images, we propose a feature category weight balance algorithm to improve the classification accuracy of minority feature categories. To validate the effectiveness of our network model for classification, we conduct experiments using Vaihingen and Postdam datasets. The experimental results demonstrate promising performance metrics: The remote sensing image classification algorithm that integrates attention mechanism and weight balance is validated in the Vaihingen dataset with pixel accuracy, mean intersection over union, and mean F1 values of 96.66%, 90.35%, and 96.66%, respectively. In the Postdam dataset, the pixel accuracy, mean intersection over union, and mean F1 values of the validated data are 95.74%, 81.47%, and 91.82%, respectively. From the classification details, incorporating an attention mechanism and a weight balance algorithm significantly enhances the recognition accuracy of cars, which account for a relatively small proportion. Specifically, the pixel accuracy of cars in Vaihingen dataset has improved by 26.44%, and in  Postdam dataset, it has increased by 21.84%, leading to commendable classification results.

Key words: attention mechanism, weight balance algorithm, DeepLabV3+network, remote sensing image, land classification