基于深度学习光谱特征提取的城市遥感地物目标分割方法

摘要:

城市遥感影像复杂背景多重语义,通过简单的形态学运算难以获取细微信息,只能获取地物目标表层特征,目标分割的像素易错乱,导致分割结果不理想。为此,提出基于深度学习光谱特征提取的城市遥感地物目标分割方法。利用高通滤波器和低通滤波器,对城市遥感图像中每个像元对应的光谱向量进行一维小波分解,应用指数函数对分解后的光谱向量进行非线性增强,凸显图像中微小光谱特征的变化。以堆栈自动编码机为基础,构建深度学习特征提取模型,将增强后的遥感图像输入模型中,即可输出图像深层次光谱特征。运用模糊C均值聚类算法,按照相似度对城市遥感图像中的所有特征像元进行划分,得到遥感地物目标初步分割结果。根据初步分割结果确定一个初始种子点,通过自适应区域生长完成地物目标的二次分割,得到修正后的分割结果。实验结果表明:面向简单背景的城市遥感图像,该方法分割结果MIoU值保持在0.55以上,而遇到复杂背景的遥感图像,其分割结果MIoU值也超过了0.5,极大提升了遥感图像分割处理质量。

Abstract:

Urban remote sensing images have complex backgrounds and multiple semantics, making it difficult to obtain subtle information through simple morphological operations. Only surface features of ground objects can be obtained, and the pixels in target segmentation are prone to confusion, resulting in unsatisfactory segmentation results. Therefore, a method for urban remote sensing land object segmentation based on deep learning spectral feature extraction is proposed. Using high pass and low-pass filters, one-dimensional wavelet decomposition is applied to the spectral vectors corresponding to each pixel in urban remote sensing images. The decomposed spectral vectors are nonlinearly enhanced using exponential functions to highlight the changes in small spectral features in the image. Based on the stack autoencoder, a deep learning feature extraction model is constructed, and the enhanced remote sensing image is input into the model to output the deep spectral features of the image. Using the fuzzy C-means clustering algorithm, all feature pixels in urban remote sensing images are divided according to similarity, and preliminary segmentation results of remote sensing land targets are obtained. Determine an initial seed point based on the preliminary segmentation results, and complete the secondary segmentation of ground targets through adaptive region growth to obtain the corrected segmentation results. The experimental results show that for urban remote sensing images with simple backgrounds, the MIoU value of the segmentation results of this method remains above 0.55, while for remote sensing images with complex backgrounds, the MIoU value of the segmentation results also exceeds 0.5, greatly improving the quality of remote sensing image segmentation processing.