| In the cold and dry areas of China,the rapid identification of river information using remote sensing technology plays a crucial role in agricultural irrigation,water resources planning and allocation,and ecological environment governance.Deep learning technology has been widely used in the identification of rivers in remote sensing images,which can efficiently and accurately extract river information,obtain real-time semantic information of rivers over a large area,and significantly reduce manpower and material costs.However,in high-altitude river valleys and snowy regions of cold and arid areas,the terrain is extremely complex,and the remote sensing river images are easily affected by factors such as ice and snow,roads,mountain shadows,and dried river beds,and there is a lack of corresponding river datasets for training deep learning networks.Therefore,the problem of identifying rivers in remote sensing images in arid regions needs to be addressed by accurately extracting river features from remote sensing images with limited labeled river data sets and obtaining high-precision semantic segmentation maps of rivers for real-time river information retrieval.This dissertation addresses the challenging issues related to river recognition in harsh and arid natural environments.It leverages the state-of-the-art image semantic segmentation network techniques and proposes innovative solutions to tackle the aforementioned challenges.Specifically,the main contributions of this study are as follows:(1)A river recognition dataset specific to the natural environment characteristics of China’s cold and dry areas was created.A large number of labeled river data are necessary for training deep learning networks to ensure that the network can effectively extract various target features from the training data.However,existing river datasets cannot meet the requirements of river recognition in cold and arid regions because they do not capture the unique geographical features of these regions.Therefore,this study collected remote sensing river images from the Google Earth and GF-2 satellites in China’s cold and dry areas,and manually annotated them with the Labelme tool.Various data augmentation methods were used to enhance the dataset.(2)AFR-Link Net network,which integrates transfer learning,residual channel attention structure,asymmetric convolution blocks,and visual activation functions,was proposed.To reduce the training time of the network,the pre-trained Res Net50 network was applied to the Link Net network using transfer learning as the encoder’s backbone network for river feature extraction.Asymmetric convolution blocks were used for model compression and acceleration,and visual activation functions were used to improve fine spatial layout.Residual channel attention structures were added to the network encoder-decoder to enhance effective segmentation features,improve the model’s segmentation ability,and obtain more detailed information.(3)A network model based on self-supervised contrastive learning was proposed.Existing network training mostly uses labeled data,but a large amount of labeled data is often difficult to obtain.Self-supervised contrastive learning can use a large amount of unlabeled data for pre-task model training,and then transfer the learned parameters to the target task.Only a small amount of labeled data is needed for downstream river extraction task fine-tuning training to achieve segmentation results superior to supervised networks.(4)A non-uniform fine-grained sampling method based on point selection strategy was proposed.Uniform sampling cannot obtain more river boundary information,and uniform sampling in all regions will increase the computational burden.The non-uniform sampling method focuses more on key data points during the sampling process,selecting representative and informative data points for sampling,and densely sampling at the river edge to obtain clear river boundary information,improve the precision of river recognition,and reduce the computational burden. |
PDF Full Text Request