摘要
为了提高红外遥感图像地物信息自动提取的精确性,同时避免人工提取遥感信息的低效性,提出了一种基于UNet深度学习模型的遥感信息提取算法。该算法用于从红外遥感图像中分割出5类地物信息(包括道路、建筑、树木、农田和水体)。首先,对分辨率高但数量较少的训练数据进行小像幅的随机裁剪,并对其进行相应的数据增强处理。然后搭建UNet深度学习模型,并用它自动提取遥感图像的特征信息。采用交叉熵损失函数以及Adam反向传播优化算法对该模型进行训练,并对测试样本中的5幅遥感图像进行精确的地物信息提取。最后,运用Jaccard指数对测试结果进行精度评定。实验结果表明,该方法对高分辨率红外遥感图像信息和可见光遥感图像信息进行了充分融合,对于不同种类地物的定位和分类都取得了较高精度。
To improve the accuracy of automatic extraction of object information in infrared remote sensing images while avoiding the inefficiency of manual extraction of remote sensing information, a remote sensing information extraction algorithm based on the UNet deep learning model is proposed. The algorithm is used to segment five kinds of object feature information including road, building, tree, farmland and water in infrared remote sensing images. Firstly, a small number of high resolution training data are cropped randomly and corresponding data enhancement processing is implemented on them. Then, a UNet deep learning model is established and is used to extract the feature information in remote sensing images automatically. The model is trained by using the cross-entropy loss function and Adam optimization algorithm and is used to extract the object information in five remote sensing images accurately. Finally, the classification result is evaluated by using the Jaccard index. The experimental results show that this method can fully fuse the high resolution infrared remote sensing image information with the visible remote sensing image information. It has higher accuracy in positioning and classification for various objects.
出处
《红外》
CAS
2017年第8期37-43,共7页
Infrared
基金
中国科学院上海技术物理研究所2015年创新专项(CX-63)
关键词
深度学习
UNet
语义分割
多光谱遥感
deep learning
UNet
semantic segmentation
multispectral remote sensing
