结合改进Laplacian能量和参数自适应双通道ULPCNN的遥感影像融合方法

Remote sensing image fusion method combining improved Laplacian energy and parameter adaptive dual-channel unit-linking pulse coupled neural network

融合SAR影像的后向散射信息和光学影像的光谱信息是提高土地覆盖分类精度的重要手段之一,其中多尺度变换是一种有效的融合方法。然而,多尺度变换方法的融合规则通常根据局部特征信息和脉冲耦合神经网络模型进行设计,存在结构信息和细节信息提取能力有限,以及脉冲耦合神经网络参数设置复杂和空间相关性差等问题。为此,本文提出一种结合改进Laplacian能量和参数自适应双通道单位连接脉冲耦合神经网络(ULPCNN)的遥感影像融合方法。该方法混合成分替换方法和多尺度变换方法,首先对多光谱影像进行IHS变换得到亮度分量I,将亮度分量I与SAR影像通过非下采样剪切波变换(NSST)分解得到高低频子带。然后对低频子带采用结合加权局部能量和八邻域修正拉普拉斯加权和的融合规则,同时对高频子带采用参数自适应双通道ULPCNN的融合规则,将高频子带的多尺度形态梯度作为链接强度,并根据OTSU阈值和影像强度来实现其他参数的自适应表示。最后依次进行NSST重建和IHS逆变换得到融合影像,并选择随机森林分类器对融合影像进行土地覆盖分类。试验结果表明,本文方法相较于13种其他方法在11个融合评价指标和土地覆盖分类精度上总体表现最佳,土地覆盖分类的总体精度和Kappa系数在区域1中比原多光谱影像分别提高了8.350%和0.107,在区域2中比原多光谱影像分别提高了6.896%和0.091。

The fusion of backscattering information of SAR images and spectral information of optical images is one of the important means to improve the accuracy of land cover classification,and multi-scale transform is an effective fusion method.However,the fusion rules of multi-scale transform method are usually designed based on local feature information and pulse coupled neural network models,and there are some problems such as limited ability to extract structural and detailed information,complex parameter settings of pulse coupled neural network and poor spatial correlation.To this end,a remote sensing image fusion method based on improved Laplacian energy and parameter adaptive dual-channel unit-linking pulse coupled neural network(ULPCNN)is proposed in this paper.This method combines the component substitution method and the multi-scale transform method.Firstly,the multi-spectral image is transformed by IHS to obtain the intensity component I,and then the intensity component I and SAR image are decomposed by non-subsampled shearlet transform(NSST)to obtain high and low frequency sub-bands.Secondly,a fusion rule combining weighted local energy and weighted sum of eight-neighborhood-based modified Laplacian is used for low frequency sub-bands,a fusion rule of the parameter adaptive dual-channel ULPCNN method is used for the high frequency sub-bands,the multi-scale morphological gradient of the high-frequency sub-band is used as the link strength,and the adaptive representation of other parameters is realized according to the OTSU threshold and image strength.Finally,the NSST inverse transform and the IHS inverse transform are performed in turn to obtain the fusion image,and the random forest classifier is selected to classify the fusion image for land cover.The experimental results show that the proposed method has the overall best performance in eleven fusion evaluation indexes and land cover classification accuracy compared with 13 other methods.The overall accuracy and Kappa coefficient of land cover classification improved by 8.350%and 0.107,respectively,in area 1,and by 6.896%and 0.091,respectively,in area 2 compared with those of the original multi-spectral images.