Dynamic Convolution Self-Attention Network for Land-Cover Classification in VHR Remote-Sensing Images

The current deep convolutional neural networks for very-high-resolution (VHR) remote-sensing image land-cover classification often suffer from two challenges. First, the feature maps extracted by network encoders based on vanilla convolution usually contain a lot of redundant information, which easily causes misclassification land cover. Moreover, these require large number parameters and high computational costs. Second, as images are complex many objects with large-scale variances, it is difficult to use popular fusion modules improve representation ability networks. To address above issues, we propose dynamic self-attention (DCSA-Net) VHR classification. proposed has advantages. On one hand, designed lightweight module (LDCM) using mechanism. This can extract more useful features than convolution, avoiding negative effect useless other context information aggregation (CIAM) ladder structure enlarge receptive field. aggregate multi-scale contexture different resolutions dense connection. Experiment results show that DCSA-Net superior state-of-the-art due higher accuracy classification, fewer parameters, lower cost. source code made public available.