Research On Remote Sensing Image Pansharpening Method Based On Deep Learning

In recent years,remote sensing satellites have provided vast amounts of remote sensing image data,these data have very important research and utilization value in many fields such as natural disaster monitoring,mineral exploration,aerospace,military,agriculture and so on.However,different remote sensing satellites use different sensors for imaging,and complex natural environmental influences the collection of remote sensing data.There are significant differences in spatial details and spectral information between acquired remote sensing images.In order to obtain multispectral images with high spatial resolution by using complex remote sensing image data,the pansharpening method was proposed and research deeply.The aim of pansharpening is to fuse the spatial and spectral information in the multisource remote sensing image data to effectively improve the spatial and spectral resolution of remote sensing images.In recent years,pansharpening methods based on deep learning have been widely and intensively studied.But there is always a slight imbalance in the retention of spectral information and the recovery of spatial information,this is also the goal of many related rearch studies.In this paper,the main work is as follows:(1)A pansharpening network(MHATNet)is proposed which is based on multiscale Transformer feature encoder modules and the hybrid attention module(HAM)with hybrid attention gate.The network mainly consists of three parts: shallow feature extraction module,multiscale Transformer feature encoder module,and the feature recovery module by the hybrid attention mechanism.The shallow feature extraction module mainly uses convolution neural networks with different convolutional kernel and CBAM mechanism,which can help us to obtain spatial and spectral information as much as possible,in this module,we expand the perceptual field by dynamically changing the convolutional kernel size.The multiscale transformer feature extraction module uses transformer block with different depth,which help us capture global information and long-range dependencies between pixel points for enhanced feature learning.The hybrid attention feature recovery module uses spatial attention and spectral attention to recover feature step by step,and we design a hybrid attention gate to balance the spatial feature and spectral feature.In the training,we use D2 SM loss to optimze out model.(2)A recurrent depth grouped attention residual pansharpening network(RGRP-Net)is proposed,which first roughly filters the feature images through convolutional layers and GELU activation function,and then acquires the residual images through a recurrent residual module which has multiple inner loops.This module consists of a depth grouped attention residual block(DGARB)and a long-short term memory convolutional update block(LCUB).Finally,the predicted image is acquired with the help of a convolutional layer.The image data input to the network is mined in depth from coarse to fine,from shallow to deep.The effectiveness of our proposed method is verified through both down-resolution and full-resolution experiments.