Land Cover Analysis Based On Multi-feature Clustering Functional Networ

Using remote sensing images for land cover analysis has broad development prospects,and the key to applying this technology is to realize the accurate segmentation of remote sensing images.At present,Convolutional Neural Networks(CNN)performs well in solving the task of image segmentation,but there is almost no semantic segmentation algorithm for land cover.When the existing semantic segmentation algorithms is applied to land cover analysis,it has the disadvantages of poor generalization and low accuracy.In order to obtain more accurate segmentation results,it is necessary to combine the characteristics of image context information and image space information,and pay attention to the intra-class differences and inter-class similarities of each prediction category.In order to achieve high-precision land cover analysis,this paper designs two models:Dual Function Feature Aggregation Network(DFFANet)and Cross Dimensional Feature Fusion Network(CDFFNet).In DFFANet,the Affinity Matrix Module(AMM)is designed to gather the context information,and the Boundary Feature Fusion Module(BFF)is designed to determine the location distribution of each category of image.In CFFANet,the Class Feature Attention(CFA)is used to gather the input features of multi-scale and multi receptive fields to capture the high-dimensional semantic information.Using the Cross Level Feature Fusion Module(CLFF)to fuse the features of different resolutions,and fuse the semantic information of different dimensions and distinguish the context correlation of each pixel.The Same Level Feature Gate Fusion Module(SLFGF)is used to dynamically fuse the information of different dimensions,filter invalid information and interference information,calibrate feature recovery process.From the experimental results of the model on the public datasets,each evaluation matric and prediction sample meet the goal.