Remote Sensing Image Processing Based On Learning Partial Differential Equations Model

Remote sensing images are very valuable in many fields such as national defense,aerospace,agriculture,and urban planning.In this age when machine learning prevails,more and more researchers pay their attention to how to combine machine learning algorithms with remote sensing image processing and how to transplant the natural image processing algorithms based on partial differential equations into remote sensing image processing.There are many issues about remote sensing image processing need to be researched and discussed.Different issue has different algorithm.This paper improves the learning partial differential equations model which is working on the natural image processing.The improved model can solve many issues about remote sensing image processing.First,this paper proposed a algorithm for cloud removal which is suitable for sparse model and can prevent over fitting.The original learning partial differential equations model is improved by establishing elastic network regression and eliminating invalid activation function.The superiority of the new algorithm had been proved by experiment.Second,aiming at the problem of spatiotemporal,the image function and image changing function in the original learning partial differential equations model is modified,and the wavelength variable is added,so that the model can take into account the characteristics of image in different spectra,and it is suitable for the processing of multi-band image.Using crank Nicolson difference to replace the central difference used in the original model,the model can obtain more overall image information.The algorithm is used to solve the problem of spatiotemporal fusion of MODIS and Landsat 8.Finally,aiming at the problem of remote sensing image classification,taking land cover classification as an example,a pixel based classification algorithm is proposed based on invariants.The original PDE learning model can't classify the whole image based on pixel,one image corresponds to one category.In this paper,one pixel corresponds to one category.This paper generalizes the dimension of differential invariants and transforms it from original two-dimensional to any multi-dimensional,thus greatly reducing the number of invariants.In the original model,the more image bands,the more invariants,the more complex the equation,and the more difficult the solution.In this paper,there are only six differential invariants in the algorithm for this problem,and the feature can still be extracted completely.Subsequent experiments show that the algorithm can achieve high classification accuracy with little sample size and time.