Research On Key Technologies Of Remote Sensing Image Land Cover Classification And Change Detection

As the key task of remote sensing image processing,the land cover classification and change detection of remote sensing image features have high practical value.In recent years,convolutional neural network has been widely used in remote sensing image processing tasks.However,remote sensing images have the characteristics of complex spatial semantic information,unbalanced categories and different resolutions,which bring new challenges to remote sensing image classification and change detection tasks.In order to solve the above problems,this paper designs and implements a multi-scale attention network which can extract the features of deep remote sensing image,and applies it to the task of remote sensing image classification and change detection.On the premise of ensuring the accuracy of the model,the model is compressed to adapt to the resource constrained platform.(1)Firstly,a multi-scale attention semantic segmentation model DIANet based on Res Ne Xt is designed and implemented for remote sensing image land cover classification task.The model is based on the encoding and decoding architecture,and uses Res Ne Xt network based on residual structure and packet convolution as the codec,which can extract the deep features of remote sensing image and overcome the problem that the network level is too deep and easy to degrade.Considering that the remote sensing image is rich in semantics and has multi-scale spatial related information,this paper embeds an improved empty space pyramid module,an improved spatial attention mechanism sis SE block and an improved channel attention mechanism cic SE into the codec of the model Block,which makes the model focus on multi-scale information of spatial level and channel level,further improves the accuracy of the model.(2)After realizing the remote sensing land cover classification model,the paper establishes the change detection model and the direct change detection model.The change detection model after classification obtains the change detection results by comparing the classification information of different phases obtained from the surface feature classification model.The direct change detection model makes use of the ability of feature classification model to extract remote sensing image features.Referring to the idea of transfer learning,the change detection results can be obtained directly by reusing the parameters of feature classification model for further training.The experimental results show that the direct change detection network is effective and can overcome the classification noise problem in the change detection model after classification.(3)In order to meet the needs of resource constrained platform,the remote sensing image land cover classification and change detection model DIANet is compressed.The depth separable convolution is used to compress the codec of the model,the region shuffle mechanism and cross spatial interaction mechanism are used to compress the spatial attention mechanism sis SE block of the model,and the cross channel interaction mechanism is used to compress the spatial attention mechanism sis SE block and channel attention mechanism cic SE of the model Block,The accuracy loss of the model is controlled within 2%,and the parameters of the model are reduced to half of the original model.