| Southwest China’s karst area is one of the largest continuously distributed karst areas in the world,accounting for 5% of the national territory and approximately 510000 square kilometers.The karst area has a wide distribution of water systems,which play a crucial role in maintaining ecological balance and human production and life in the karst area.Therefore,quickly and accurately obtaining water body information in karst areas is of great significance and role in water ecological protection,dynamic monitoring of water body changes,water conservancy planning,flood monitoring,and other aspects.Inspired by computer vision research in recent years,convolutional neural networks have been continuously proposed for remote sensing image analysis tasks.However,due to the large number of mountain shadows in karst areas,water bodies have the characteristics of small inter class variance and large intra class variance,resulting in low segmentation accuracy.Therefore,the work of this article will revolve around how to improve the above issues,and the specific content is as follows:(1)A water extraction network SA-DeNN based on improved strip pooling and attention mechanism has been developed.The model is based on the DeepLabv3+model structure,with ResNet101 as the backbone network.Firstly,a Mixed Pooling Module(MPM)improved by both Strip Pooling(SP)and Pyramid Pooling(PP)is added to the Atrus Spatial Pyramid Pooling(ASPP)module,which can aggregate information of different types of water bodies and meet the segmentation requirements of water bodies with varying sizes in remote sensing images.Additionally,thanks to the elongated and narrow shape of the strip pond,it can capture water bodies that are widely present in remote sensing images,reduce the impact of mountain shadows,and improve the segmentation accuracy of water bodies;Secondly,in order to further enhance the model’s ability to focus on key information such as water bodies,an Effective Channel Attention(ECA)mechanism is introduced in the encoder section of the model.By utilizing the attention mechanism,the model can pay more attention to the advantages of important information based on feature weights,further enhancing the model’s ability to learn features;Finally,through ablation experiments and experimental comparison with other classic semantic segmentation models on two datasets,the experimental results show that compared with the DeepLabv3+model,the overall accuracy(OA),recall(R)and mean intersection over Union(MIoU)of the improved SA DeNN model on the Sentinel-2dataset and Google Earth dataset have increased by 2.76%,2.79% 2.64% and 2.01%,2.53%and 2.92%,respectively,demonstrate the effectiveness of model improvement,and SA DeNN exhibits better extraction performance compared to other semantic segmentation models.(2)A water body extraction network SD-DeNN based on multi-scale dense connected feature extraction is proposed.First,we replace the pooling of empty space pyramids in DeepLabv3+with the pooling of dense empty space pyramids DenseASPP,and also introduce the bar pooling SP,which can extract multi-scale information and obtain a sufficiently large Receptive field under the limited void rate,further improving the ability of pattern learning features.Secondly,ResNet34 with residual learning module will be pre trained as the backbone network of the model to further reduce the number of parameters and time required for model training.Finally,through experimental comparison with other semantic segmentation models on two datasets,the experimental results show that compared with the basic DeepLabv3+model,the three main indicators of the SD DeNN model on Sentinel-2 remote sensing image dataset and Google Earth remote sensing image dataset,OA,R,and MIoU,respectively,have increased by 2.23%,2.58%,2.12% and 1.98%,1.24%,1.65%.The effectiveness of this model in extracting water bodies from remote sensing images of karst landforms has been demonstrated. |
PDF Full Text Request