| Optical remote sensing images have high resolution,large scale,periodic acquisition and other characteristics.Optical remote sensing images can record rich information of terrain texture,geometry,spectrum and other features.Road extraction technology for remote sensing images has important significance for urban planning,intelligent navigation,disaster monitoring and other fields.In recent years,road extraction methods based on convolutional neural networks have made significant progress.However,these methods still face problems such as weak long-distance dependency modeling ability and insufficient robustness to cloud-occlusion.Self-attention deep model and cloud-occlusion robust method for road extraction from remote sensing images was presented to address the challenges in road extraction from remote sensing images.The main contributions of this thesis are given as follows:(1)A road extraction method from remote sensing image based on self-attention mechanism was proposed.To address the problem of weak long-distance dependency modeling ability of convolutional operations and imbalanced categories in remote sensing images,this thesis proposed a road extraction method based on self-attention mechanism using a mixed loss function.Firstly,residual units were constructed based on selfattention mechanism to enhance the modeling ability of long-distance dependencies.Secondly,a model optimization was established based on cross-entropy loss function and Dice coefficient loss function to avoid the problem of imbalanced categories of road pixels being ignored by the model while focusing too much on background pixels with larger quantities.(2)A robust road extraction method for cloud-occlusion based on invariant regular constraints was proposed.To address the problem of insufficient robustness to cloudocclusion in road extraction from remote sensing images,this thesis proposed a robustness to cloud-occlusion road extraction method based on invariant regularization optimization.This method used feature invariant loss function and classification invariant loss function to constrain the model to have similar feature vectors and classification results between cloud-covered images and corresponding cloud-free images,thereby improving the robustness to cloud-occlusion of road extraction from remote sensing images. |
