低空无人机遥感影像的语义分割与地物分类模型优化研究

Research on Model Optimization for Semantic Segmentation and Land Cover Classification of Low-Altitude UAV Remote Sensing Images

摘要: 随着低空无人机遥感技术的快速发展，其所获取到的高分辨率影像为开展精细化地物的识别以及分类，提供了丰富的数据源。然而，当前基于深度学习来开展的语义分割模型在处理这类影像时，经常会面临地物边界模糊、类内差异大以及类间相似性高等方面的挑战，并且大多数研究仅仅把RGB颜色信息当作信息来使用，由于没有充分挖掘影像多维光谱特性，导致模型在复杂场景当中的分类精准度以及鲁棒性方面是有提高空间的。为此，本研究把重点聚焦在有效开展RGB空间信息以及多光谱特征的融合工作，对面向低空无人机遥感影像的语义分割以及地物分类模型进行优化，搭建双分支特征提取网络，分别去处理RGB图像以及特定变换后的光谱信息，设计跨模态特征开展交互与融合的模块，来实现空间细节和光谱判别特征的互补增强。实验结果显示，所提出的融合模型在建筑、植被、道路、水体等多个典型地物类别分类精度方面，明显优于仅仅运用RGB信息的基准模型，尤其在区分光谱特征相似但空间纹理不同的地物方面表现突出，有效改善了分割边界的连续性与分类结果的同质性。本研究不仅为无人机遥感影像的自动化解译提供了一种精度更高、适应性更强的技术方案，而且通过探索多源信息融合机制，对推动精准农业、城市测绘、环境监测等领域的智能化应用具有重要的实践参考价值。

Abstract: With the rapid development of low-altitude unmanned aerial vehicle (UAV) remote sensing technology, the high-resolution images obtained provide abundant data sources for the identification and classification of refined land cover features. However, current deep learning-based semantic segmentation models often face challenges such as blurred land cover boundaries, large intra-class variations, and high inter-class similarity when processing such images. Moreover, most studies only employ RGB color information as input, and the insufficient exploitation of multi-dimensional spectral characteristics of images leaves room for improvement in classification accuracy and robustness of the models in complex scenarios. To address these issues, this study focuses on the effective fusion of RGB spatial information and multi-spectral features, and optimizes semantic segmentation and land cover classification models for low-altitude UAV remote sensing images. A dual-branch feature extraction network is constructed to process RGB images and spectrally transformed information separately, and a cross-modal feature interaction and fusion module is designed to achieve complementary enhancement of spatial details and spectral discriminative features. Experimental results demonstrate that the proposed fusion model significantly outperforms baseline models using only RGB information in classification accuracy for multiple typical land cover categories including buildings, vegetation, roads, and water bodies. It exhibits outstanding performance especially in distinguishing land cover features with similar spectral characteristics but different spatial textures, effectively improving the continuity of segmentation boundaries and the homogeneity of classification results. This study not only provides a technical solution with higher accuracy and stronger adaptability for the automated interpretation of UAV remote sensing images, but also offers important practical reference for promoting intelligent applications in precision agriculture, urban mapping, environmental monitoring and other fields by exploring multi-source information fusion mechanisms.