Research On Segmentation Method Of Environmental Remote Sensing Image Based On Deep Learning

With recent rapid development of the space remote sensing technology,surface natural environment information can be collected by environmental remote sensing satellites,and large-scale environmental remote sensing monitoring data is available for different part of the earth.How to efficiently identify and accurately interpret the natural rivers and forests from the multi-source massive heterogeneous environmental remote sensing data,and achieve large-scale and global surface natural object monitoring,is an important mission within the environmental remote society,and show high value for the sustainable development of the natural environment.Toward this task,researchers have applied deep learning technic on remote sensing data analysis,but the current research has not fully considered the structural characteristics of massive heterogeneous remote sensing data,and the research is still in the initial stage.The main challenge is how to combing the limited ground truth information,the different type of the data feature,the difference and similarity between each type of the remote sensing data into the deep learning based remote sensing feature extraction model.This dissertation carries out research on deep feature learning based environmental remote sensing analysis model,aiming to provide new method for environmental remote sensing analysis,in order to promote the development and application of deep learning in environmental remote sensing.The main research contents are as follows:1.Aiming at the problem that deep learning model training is not easy to converge with limited sample space,and the limitation of existing semi-supervised learning model for the use of unlabeled high-dimensional remote sensing data information,a few-shot learning based remote sensing image segmentation method is proposed.This method constructs a segmentation model architecture for limited labeled Remote sensing image samples with few-shot learning,and uses the proposed unsupervised learning method for the unlabeled remote sensing image feature extraction.Then,a knowledge transformation based model pre-training method is proposed for the model pre-training.Finally,a semi-supervised learning based feature extraction model is used for optimizing the image segmentation model in limited sample space by combining the unlabeled image information and a small amount of labeled image information.The experiment shows that,compared with the state-of-the-art deep learning based image segmentation methods,this method can better describe the features of environmental remote sensing images in small sample datasets with limited ground truth data,and obtain better forest object segmentation performance.2.Aiming at the different characteristics for the surface natural object among different time and location,and the limitation of current deep learning model for the environmental remote sensing object semantic feature extraction,a semantic fusion based large scale remote sensing data segmentation method is proposed.This method first divided the samples into different ground based on the proposed enhanced fully convolutional semantic feature extraction model.Then,a semantic based distance measurement metthod is proposed.Finally,the output of multi semantic segmentation model is fused based on the proposed semantic fusion method,and provide the finally segmentation result.The experiment shows that,compared with the state-of-the-art semantic segmentation methods,the proposed method could better describe the different of the large scale environmental remote sensing river data,and achieve better river segmentation result.3.Aiming at the different structure of the multi-source remote sensing data,and the limitation of current deep learning model for the multi-source data analysis,a multi-scale feature based multi-source remote sensing data segmentation method is proposed.This method first provides a multi-scale feature extraction model.Then,this method proposed an multi feature fusion structure for multi-scale feature information processing.In the meantime,an improved residual network model is adopted to characterize the high-dimensional feature information of remote sensing image data.The experiment shows that,compared with the state-of-the-art segmentation methods,the proposed method could better extract the multi-scale feature,and achieve better river segmentation result.4.Aiming at the common information of the remote sensing data among different resources,and the current limitation of the contrastive learning for the remote sensing feature description,a contrastive learning based multi-source remote sensing data segmentation method is proposed.This method first extracts the common information between multi-source remote sensing data,based on the proposed contrastive learning model.Then,the self-feature of each remote sensing data is extracted based on the proposed self-information extraction method.Finally,the common and self-information is added into the multi-source remote sensing segmentation model,along with the multi-source remote sensing data information.The experiment shows that,compared with the state-of-the-art segmentation methods,the common information between multi-source remote sensing data could better help the segmentation process,and achieve better river segmentation result from the remote sensing data.