MDMASNet: A dual-task interactive semi-supervised remote sensing image segmentation method

•A MDMF module is proposed to enhance the extraction of variable size and shape target features by the network.•Mixed attention is proposed for aggregating contextual information over long distances, reducing the possibility of classification errors.•To accelerate network convergence and prevent network overfitting, we use residual connection to solve network degradation caused by network over depth.•We use SE attention to allocate the weights of different channels to strengthen the discriminator's evaluation ability of unlabeled segmentation results.