Resumen
Semantic segmentation algorithms leveraging deep convolutional neural networks often encounter challenges due to their extensive parameters, high computational complexity, and slow execution. To address these issues, we introduce a semantic segmentation network model emphasizing the rapid generation of redundant features and multi-level spatial aggregation. This model applies cost-efficient linear transformations instead of standard convolution operations during feature map generation, effectively managing memory usage and reducing computational complexity. To enhance the feature maps? representation ability post-linear transformation, a specifically designed dual-attention mechanism is implemented, enhancing the model?s capacity for semantic understanding of both local and global image information. Moreover, the model integrates sparse self-attention with multi-scale contextual strategies, effectively combining features across different scales and spatial extents. This approach optimizes computational efficiency and retains crucial information, enabling precise and quick image segmentation. To assess the model?s segmentation performance, we conducted experiments in Changge City, Henan Province, using datasets such as LoveDA, PASCAL VOC, LandCoverNet, and DroneDeploy. These experiments demonstrated the model?s outstanding performance on public remote sensing datasets, significantly reducing the parameter count and computational complexity while maintaining high accuracy in segmentation tasks. This advancement offers substantial technical benefits for applications in agriculture and forestry, including land cover classification and crop health monitoring, thereby underscoring the model?s potential to support these critical sectors effectively.