基于深度学习与数据增强的土地利用分类研究

Research on land use classification based on deep learning and data enhancement

关键词:  卷积神经网络(CNN)  深度学习  土地利用分类  珠海一号  U-Net

Key words:  CNN  Deep learning models  Land use classification  Zhuhai-1  U-Net

摘要:

【目的】文章旨在探讨如何利用深度学习技术提高高光谱图像分类的准确性，解决高光谱数据处理中面临的维度高、信息冗余及融合过程中的信息损失等挑战，为土地利用管理、环境监测及城市规划等提供强有力的技术支撑。【方法】以象山港为研究区域，基于“珠海一号”高分辨率遥感影像，结合主成分分析（PCA）与信息损失引导图像融合法（ILGIF）作为关键数据增强技术，深入开展了土地利用分类研究。通过PCA方法，有效降低了影像数据的维度，提取了最具代表性的特征，同时去除了噪声和冗余信息。而ILGIF技术则在数据降尺度的过程中，通过最小化信息损失来优化图像融合，保留了关键的空间和光谱信息，为分类任务提供了更高质量的数据支持。在此基础上，采用深度学习中的U-net模型，对土地利用类型进行了精确分类。【结果】相较于传统方法，采用深度学习中的语义分割方法能够有效提高精度水平，U-net模型的整体分类精度为90%，研究区内水体、林地、建成区的分类精度良好，依次为96.93%、88.34%和71.69%。【结论】后续将进一步挖掘多模态（光学与微波）数据资源的潜力，融合多尺度的空间与光谱分辨率数据集的优势，探索并推动人工智能、大数据等技术集成在遥感图像分类领域的创新应用，实现对地表覆盖快速且精确的监测，为相关部门提供数据支撑。

Abstract:

[Purpose] The purpose of this paper is to explore how to use deep learning technology to improve the accuracy of hyperspectral image classification, solve the challenges of high dimensionality, information redundancy and information loss in the process of hyperspectral data processing, and provide strong technical support for land use management, environmental monitoring and urban planning. [Method] Taking Xiangshan Port as the research area, based on the high-resolution remote sensing image of Zhuhai-1, combined with principal component analysis (PCA) and information loss-guided image fusion method (ILGIF) as the key data augmentation technology, the research on land use classification was carried out in depth. Through the PCA method, the dimension of the image data is effectively reduced, the most representative features are extracted, and the noise and redundant information are removed. In the process of data downscaling, ILGIF technology optimizes image fusion by minimizing information loss, retains key spatial and spectral information, and provides higher quality data support for classification tasks. On this basis, the U-net model in deep learning is used to accurately classify land use types. [Result] Compared with the traditional method, the semantic segmentation method in deep learning can effectively improve the accuracy level, and the overall classification accuracy of the U-net model is 90%, and the classification accuracy of water, woodland and built-up area in the study area is good, which is 96.93%, 88.34% and 71.69%, respectively. [Conclusion] In the future, we will further explore the potential of multimodal (optical and microwave) data resources, integrate the advantages of multi-scale spatial and spectral resolution datasets, explore and promote the innovative application of artificial intelligence, big data and other technologies in the field of remote sensing image classification, realize fast and accurate monitoring of land surface coverage, and provide data support for relevant departments.