Buildings Extraction from Remote Sensing Data Using Deep Learning Method Based on Improved U-Net Network
Buildings Extraction from Remote Sensing Data Using Deep Learning Method Based on Improved U-Net Network
复制标题
基于改进U-Net网络的深度学习方法遥感数据建筑物提取
DOI：
10.1109/igarss.2019.8899798
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
Lin Sun
中科院分区：
文献类型：
--
作者：
Yiru Duan;Lin Sun
由于建筑物的形状各异，且与周围各种地物类型交叉分布，传统的分类方法难以高精度地提取建筑物。基于神经网络的深度学习方法可以深入挖掘遥感图像的有用信息，提高建筑物识别的准确率。但由于神经网络涉及的参数多，对训练样本的需求量大，使其在建筑物提取中的应用受到限制。为了提高深度学习方法在遥感图像建筑物提取中的精度，在U网网络中插入身份跳跃连接进行样本训练，有效减少了参数数量，显著减小了模型规模，避免了层数加深导致的梯度爆炸，明显提高了分割精度。通过对不同层位计算结果的比较，表明随着层位的加深，计算精度逐渐提高。
Due to the different shapes of buildings and the cross-distribution with various surface types around them, it is difficult to extract buildings in high precision using traditional classification methods. The deep learning method based on neural network can mine useful information of remote sensing image in depth and improve the accuracy of building recognition. However, the application of neural network in building extraction is limited because of the large number of parameters involved and the large demand for training samples. In order to improve the accuracy of building extraction in remote sensing images by using deep learning method, identity skip connection is inserted into U-net network for samples training, which effectively reduces the number of parameters, significantly reduces the size of the model, and avoids the gradient explosion caused by the deepening of the number of layers, and obviously improves the accuracy of segmentation. By comparing the results of different layers, it is shown that with the deepening of layers, the accuracy increases.