Research On Land Use Classification Of High-Resolution Image Based On Multi-Scale Information Fusion

In recent years,remote sensing image land use classification technology has developed rapidly,and people can obtain a large number of land classification results of high-resolution remote sensing images with different spatial,spectral and temporal resolutions,which provides important information sources and analysis conditions for various practical applications such as urban planning,geological investigation and environmental assessment.However,remote sensing images are more complex than natural images,and multi-class parcels often have different scale,spectral and texture information.It is difficult for existing methods to express multi-class parcels using highly generalized semantic features,and the parcels in the classification results often have incomplete structure and inaccurate edge delineation.To address the above problems,this paper is based on the idea of multi-scale information fusion,combining multi-scale feature extraction and fusion,attention mechanism and super pixel image segmentation,with the goal of improving the classification accuracy of land use multiclass parcels in the following two parts.(1)High resolution image land use classification method based on SMH-Net.Aiming at the problems of remote sensing image land use classification,there are a lot of complex topography and spectral confusion in the image which easily cause blurred parcel outline and misclassification,this paper proposes SMH-Net land use classification method.The method mainly carries out the following works: firstly,it adopts dense residual module and cascade feature fusion method to construct multiscale feature extraction network MH-Net,which fully extracts multi-class parcel semantic information and fuses multi-class features to improve the extraction capability of multi-class parcel features in complex scenes;secondly,it combines differentiable SLIC algorithm to carry out deep clustering with the extracted multi-class parcel feature information to form accurate edge segmentation Third,the adaptive weighted feature fusion strategy is designed for the multi-scale feature part,and the weight vector is updated adaptively with the help of network learning,so as to make better use of different layers of features.(2)High resolution image land use classification method based on MLUM-Net.Aiming at the problem that parcels in remote sensing images often have multiscale shape features,which leads to incomplete extraction of parcel structure and noise,this paper introduces the idea of multi-scale feature learning to design the MLUM-Net land use classification method.Firstly,the MDSPA module is designed as an encoder to build the network down-sampling process by combining multi-scale hole convolution and hybrid spatial pyramidal attention mechanism to improve the multiscale feature extraction capability and the accuracy of parcel location;secondly,the HPP optimization module is proposed to avoid the semantic loss caused by upsampling and improve the information flow to obtain rich global contextual information through multi-scale pooling,enhance the multi Finally,in order to solve the problem of unbalanced percentage of categories in land use datasets,we design a hybrid loss function by combining the characteristics of structural diversity of multi-class parcels,balance the attention of the network to different parcel types,and enhance the optimization of noise areas and wrong classification results to further refine the classification results.The methods proposed in this paper can improve the accuracy of land use classification of high-resolution images to a certain extent.Compared with the commonly used segmentation networks,the classification performance indexes are significantly improved,which is of great significance for the future research and practical application of land use classification methods.