Study Of Remote Sensing Image Segmentation And Classification Algorithm Based On Learning Method

With the development of remote sensing technology,computer technology and big data technology,the automatic analysis technology of remote sensing image has become an important mean for people to acquire geospatial information.It has important application value in the fields of military reconnaissance,urban planning,resource utilization and environmental monitoring.We deeply studies the remote sensing image segmentation and classification algorithm based on learning method.This paper mainly includes two aspects,remote sensing image river segmentation algorithm based on machine learning and multi target classification algorithm based on deep learning.In this paper,remote sensing image river segmentation algorithm based on multi-feature fusion and soft voting is deeply studied.Aiming at the problem that the single feature can not effectively distinguish the river and the background area,resulting in low recognition rate of the classification model,this paper proposes a feature extraction method based on the local entropy,texture and spectral information of the remote sensing image.After normalizing the features,the classifier is used to identify the results,so that the coarse segmentation results can be obtained.Aiming at the misclassified pixels in the rough segmentation results of river,this paper proposes an optimization method based on morphological processing combined with multiple criteria soft voting.We uses open and close operations to correct misclassified pixels and uses similarity rules consisting of color and texture similarity,and corners distribution to eliminate interference from difficult samples.Aiming at the problem that the river segmentation profile is not close enough to the river bank line,an optimization algorithm based on level set geometric active contour is proposed to approximate the river bank line.It processes the river contour into an evolving activity curve based on the evolution principle of active contour curve.When the overall energy is at a minimum,the result of river segmentation result close to the riverbank is obtained.Experiments show that the proposed algorithm can accurately segment river region in complex backgrounds.In this paper,the multi target classification algorithm of remote sensing image based on convolutional neural network is deeply studied.After analyzing the design principles and the existing problems of UNet and SegNet,this paper proposes an improved remote sensing image classification network,which combining the advantages of both and making innovations.In order to fully exploit the different levels of information in remote sensing images,this paper designs a multi-scale feature coding structure,and designs a decoding structure,which combines the underlying information of the image and the high-level semantic information.In order to improve the recognition rate of the target edge and its position,,the maximum pooled index upsampling layer is designed.In order to improve the recognition rate of targets with fewer pixels such as roads in remote sensing images,a cost sensitive loss function is designed.It improves the robustness of the network model by increasing the penalty for misclassification of this type of target during the training process.We add a batch normalization acceleration layer to the network to speed up the network training.Experiments show that compared with other networks,the improved network has achieved better results in terms of accuracy,target detail recognition and consistency of prediction results.In summary,this paper makes full use of the advantages of machine learning and deep learning in dealing with pattern recognition problems,and proposes an accurate river segmentation and multi-target classification algorithm for remote sensing images.Experiments show that the algorithm is effective and has certain theoretical significance and application value.