236 / 2023-09-29 15:28:26

多尺度信息融合的遥感影像建筑物提取方法

多尺度信息；建筑物提取；语义分割；注意力机制；空洞卷积

在大量且复杂的遥感影像中提取建筑物信息一直是遥感智能应用的重要研究内容。针对复杂环境下的遥感影像建筑物分割不精准及小型建筑物分割易被忽略等问题，本文提出了一种基于注意力机制和Deeplabv3+的遥感影像语义分割算法—SC-deep网络。该网络采用编码—解码结构，利用主干残差注意力网络提取深层特征和浅层特征，通过空洞空间金字塔池化模块和通道空间注意力模块聚合遥感影像的空间和通道信息权重，有效利用了遥感影像建筑物的多尺度信息，从而减少影像细节在训练中的损失，提高模型的提取精度。实验结果表明，本文方法在Aerial imagery dataset数据集上的分割结果均优于其他主流分割网络，能够有效识别并提取复杂建筑物边缘和小型建筑物，表现出更优异的建筑物提取性能。

Extracting building information from a large number of complex remote sensing images has always been an important research content of remote sensing intelligent applications.Aiming at the problems of inaccurate building segmentation in remote sensing images in complex environments and easy neglect of small building segmentation, this paper proposes a semantic segmentation algorithm for remote sensing images based on attention mechanism and Deeplabv3+ network-SC-deep. The network adopts the encoding-decoding structure, extracts deep features and shallow features by using the backbone residual attention network, aggregates the spatial and channel information weights of remote sensing images through the hollow space pyramid pooling module and channel spatial attention module, and effectively uses the multi-scale information of buildings in remote sensing images, thereby reducing the loss of image details in training and improving the extraction accuracy of the model. Experimental results show that the segmentation results of the proposed method on the Aerial imagery dataset are better than other mainstream segmentation networks, and can effectively identify and extract complex building edges and small buildings, showing better building extraction performance.