Shadow Enhancement And Classification Of Hyperspectral Image Based On Stochastic Resonance And Convolution Neural Network

Hyperspectral images(HSIs)with rich spatial-spectral information are collected by high-resolution remote sensing equipment,and are widely used in fields like environmental detection,urban planning,and military.The classification and information extraction of ground objects in HSIs also become an important research direction.However,shadow regions with low feature expression capability are usually contained,affecting the abundant information exploration and extraction in classification.Besides,the commonly used convolutional neural network(CNN)for HSIs classification ignores the importance distinction of key and secondary features,leading to insufficient feature utilization,and attention mechanism(AM)is found as an embedded structure that can automatically learn the contribution of input to output,which is potential and introduced to optimize the function.Therefore,the improvement of the information expression ability of the shadow region is focused on in this paper,and the two-dimensional(2D)dynamic stochastic resonance(DSR)iterative equations are derived for more spatial correlation utilization,thus promoting feature extraction in classification.Besides,AM is introduced to improve the function of CNN.And a multi-attention module(MAM)is raised to extract the channel and spatial attention information by independently selecting important features.The main content of this article is as follows:(1)Based on the principle of 1D DSR in signal enhancement by utilizing noise,an iterative 2D DSR equation that can better utilize spatial information is derived for HSIs processing.The information in the shadow region is enhanced by making use of information in the neighborhood in four directions up,down,left,and right.Experiments are conducted on real HSI datasets with shadows.Compared to the original data,2DDSR has a significant effect on promoting the brightness and spectral curve of the image.(2)AM is introduced into CNN for attention extraction and classification effect improvement.Several applications and combinations of multiple attention modules are discussed,and a muti-attention module is proposed for spatial-channel attention extraction,which is composed of convolutional block attention module(CBAM)and efficient channel attention(ECA).The overall accuracy of HYDICE,Indian Pines,and Pavia University datasets reach 96.7585%,99.4268%,and 99.9445% respectively,which increased by 0.445%,0.4512%,and 0.2689% compared to before.(3)To make better use of signals in HSIs shadow region,1D and 2D DSR are adopted for shadow enhancement in spectral and spatial dimensions respectively,and the MAM-3DCNN is applied to the enhanced data.The overall accuracy of HYDICE is97.75%,which is 1.44% higher than that before,which shows that DSR’s spectralspatial processing can further promote information expression,and has better results compared to current commonly used classification networks.