结合深度学习技术的遥感影像城市建筑物自动提取

Automatic Extraction of Urban Buildings from Remote Sensing Images Combined with Deep Learning Technology

摘要: 背景：目前建筑物提取的自动化程度仍然不够高，原有的机器学习算法无法满足建筑物提取的需求。深度学习技术的兴起为信息提取提供了新的方法。本文目的是研究遥感影像城市建筑物的自动提取，以达到近似人工矢量提取的效果，方法是以深度学习技术为基础，研究1 m空间分辨率数据，通过分辨率增强，得到0.5 m、0.25 m空间分辨率数据，采用DeepLab v2深度卷积神经网络，进行实验分析，并通过矢量简化的方法，研究提取结果的后期优化，以符合对建筑物提取的要求。结果发现通过分辨率增强方法，像素精度提高了7.9%，交并比提高了11.8%，在此基础上，结合矢量优化方法，对提取的建筑物边缘进行优化，结果更加符合建筑物提取需求。结论是针对深度学习技术特点，进行分辨率增强可以提高信息提取效果，结合矢量方法进行优化，可以改善建筑物提取效果。研究应用意义及改进方向：将深度学习技术应用于遥感信息提取，可使遥感信息提取的精度更高。未来将对遥感影像特征进行更深入的优化研究。

Abstract: Background: At present, the level of automation of building extraction is still not high enough. The original machine learning algorithms cannot meet the needs of building extraction. The rising of deep learning technology provides new methods for information extraction. Objective: Study the automatic extraction of remote sensing image urban buildings to achieve the effect of approximate artificial vector extraction. Methods: The DeepLab v2 deep convolutional neural network is used for experimental analysis, and the vector optimization method is used to study the later optimization of the extraction results to meet the requirements for building extraction. Experimental content: The original data is subjected to resolution enhancement of the original data to obtain spatial resolution data of 0.5 m and 0.25 m. Results: Over-resolution enhancement method improves pixel accuracy by 7.9% and increases cross-ratio by 11.8%. Based on this, combined with vector optimization method, the extracted building edges are optimized, and the results are more in line with building extraction requirements. Conclusion: For the characteristics of deep learning technology, the resolution enhancement can improve the information extraction effect and the vector method can be used to optimize the building extraction effect. Research application significance and direction of improvement: Applying deep learning technology to remote sensing information extraction, the accuracy of remote sensing information extraction is higher. Deeper optimization research on the characteristics of remote sensing image will be carried out in the future.