Hyperspectral Image Reconstruction Based On Multi-scale Analysis And BP Neural Network

Because of its rich spectral information and spatial information,Hyper-Spectral(HS)images are widely used in precision agriculture,military reconnaissance,ground feature exploration,and disaster prediction.However,due to the hardware limitations of current sensor devices,HS images have HS resolution,but their spatial resolution is usually lower.Therefore,super-resolution reconstruction technology is used to improve the spatial resolution of HS images.Most of the current HS image spatial resolution reconstruction work is aimed at HS images within the multi-spectral range,and relatively few reconstruction work for HS images outside the MS band range.In order to realize the super-resolution reconstruction of HS images in the full-band range,this paper proposes a super-resolution reconstruction method of HS images that combines multi-scale analysis and BP neural network.The main research contents include:(1)The characteristics of the spectral bands of the Multi-Spectral(MS)image and the HS image are analyzed,and the bands of the HS image are divided.Among them,the bands whose spectral range is the same as the spectral range of the MS image are regarded as matching bands,and the bands whose spectral range is different from that of the MS image are regarded as non-matching bands.(2)HS image matching band reconstruction based on wavelet transform.Firstly,the selection principle of weighted average is used to fuse the low-frequency wavelet coefficients of HS image and MS image,and then the selection principle of maximum area energy is used to fuse the high-frequency wavelet coefficients of HS image and MS image.The fused coefficients are subjected to inverse wavelet transform to obtain reconstructed images of matching bands.(3)Compared with the wavelet transform method,the second generation Curvelet transform method is used to reconstruct the matching band of the HS image.The selection principle of the low-frequency coefficients is weighted average,and the selection principle of the high-frequency coefficients is that the absolute value is the largest.The fused coefficients are subjected to inverse Curvelet transform to obtain the reconstructed image of matching bands.(4)Reconstruction of unmatched bands of HS images based on BP neural network.This method uses the before and after changes of pixel values between the reconstruction results of matching bands and the original HS image as training knowledge to construct a BP neural network and conduct network training.The pixels of the unmatched band of HS image are used as the input of the network to simulate the experiment.Then,the simulation results are optimized by combining the normalized information entropy of HS image and the neighborhood spatial relationship of pixels.In this paper,the effectiveness of the algorithm is verified by taking sentinel-2,Gaofen-1and Gaofen-5 data as examples,and the fusion results are compared with the simulation results.The experimental results show that the proposed method combining multi-scale analysis and BP neural network can achieve the full band super-resolution reconstruction of HS image,and has a good reconstruction effect.The paper has 42 diagrams,6 tables and 62 references.