Research On Two-step Strategy Subpixel Mapping Of Remote Sensing Images

Due to the limitation of sensors,environment,and heterogeneous of land cover types,many remote sensing images contain mixed pixels.Mixed pixel is a pixel contains more than one land cover types.Using the traditional pixel-based hard classification,mixed pixels would be misclassified.Soft classification can only generate the portion of different land cover types in mixed pixels.The location of the land cover types in mixed pixel is absent.Subpixel mapping technique classifies images in subpixel scale,and can use the output of soft classification to locate the mixture of different land cover types in mixed pixels.The mixed pixel issues in remote sensing images can be handled.Meanwhile,the subpixel mapping can enhance the spatial resolution of land cover image.The spatial resolution of the output of subpixel mapping is finer than the input.Based on the previous study of subpixel mapping,the content of our study are as follows.1.A two-step strategy based subpixel mapping method is used.Moreover,a simple and efficient soft classification method is used to produce fraction images.Results indicate that,compared with the previous methods,the proposed methods can generate the most accurate land cover maps on various remote sensing data set.2.The soft value estimation of two step strategy is similar with the image super resolution.Therefore,three different image super resolution neural network-based subpixel mapping methods are proposed,Besides,the efficient subpixel convolutional neural network-based subpixel mapping(ESPCN-SPM)uses a transfer learning strategy to update the network parameters to reduce the time consumption and overcome the limitation of training images.Therefore,the time consumption can be reduced remarkedly.The proposed method can generate a higher accuracy result than the others with a similar time consumption of pixel swapping algorithm(PSA).3.The accuracy of subpixel mapping can be influenced by scale factors,complexity of land cover types and the error of soft classification.In order to overcome the aforementioned limitation,a deeper architecture is used to estimate soft class value in the proposed deep image prior-based subpixel mapping(DIP-SPM).Meanwhile,the deep image prior(DIP)strategy is used to update the network parameters without any assistant of training date.The experiment demonstrates that the accuracy of the proposed method decreases steadily with the scale factor increasing,the complexity of land cover types increasing,and soft classification error increasing.Meanwhile,the proposed method can generate the most accurate results.