基于深度主题模型的高分辨率遥感影像场景分类方法研究

High-resolution remote sensing image scene classification is an effective way to bridge the semantic gap between the low-level features and high-level semantic information. Scene classification methods based on mid-level features and deep learning are two representative methods for high-resolution remote sensing image scene classification. Due to the complexity and variability of high-resolution remote sensing images, scene classification methods based on mid-level features have limited model fitting ability and feature discriminative ability. Scene classification method based on deep learning usually relies on training samples and cannot effectively extract significant local features in the scene. This project will combine the flexible implicit semantic mining ability of the topic model, and the powerful feature self-learning ability and nonlinear fitting ability of deep learning, and propose the research on high-resolution remote sensing image scene classification method based on deep topic model. The core idea is based on the main line from multi-scale homogeneous-heterogeneous low-level feature learning and mid-level supervised topic modeling based on deep fitting, to high-level deep topic joint semantic understanding. This project will construct the supervised topic model based on deep fitting, to optimize the parameters of topic modeling, using the homogeneous-heterogeneous visual representation. The deep features are mined from the semi-supervised convolutional neural networks, and are then combined with topic features for adaptive high-level semantic interpretation. This study can improve the application potential of high-resolution remote sensing images, therefore has important theoretical and applied significance.

高分辨率遥感影像场景分类是跨越底层特征和高层语义信息之间语义鸿沟的有效途径。场景分类代表性方法包括基于中层特征的场景分类方法和基于深度学习的场景分类方法，由于高分辨率遥感影像的复杂多变性，基于中层特征的场景分类方法往往中层特征区分性及模型拟合能力不足，基于深度学习的场景分类方法依赖大数据样本且无法有效挖掘场景中代表性局部特征。为此，本项目拟综合主题模型灵活的隐含语义挖掘能力和深度学习强大的特征自学习及非线性拟合能力，开展基于深度主题模型的高分辨率遥感影像场景分类方法研究。核心思想是以“多尺度同异质底层特征学习—主题特征深度拟合建模—深度主题特征融合的场景语义理解”为主线，基于同异质视觉表征，构建主题特征深度拟合的监督主题模型，实现模型参数推理优化；基于半监督卷积神经网络挖掘深度特征，并融合主题特征进行自适应场景语义分类，提升高分辨率遥感影像的应用潜力，具有重要的理论与应用意义。

由于高分辨率遥感影像的复杂多变性，基于中层特征的场景分类方法往往中层特征区分性及模型拟合能力不足，基于深度学习的场景分类方法依赖大数据样本且无法有效挖掘场景中代表性局部特征。本项目主要综合主题模型灵活的隐含语义挖掘能力和深度学习强大的特征自学习及非线性拟合能力，开展基于深度主题模型的高分辨率遥感影像场景分类方法研究。本项目攻克了多尺度同异质底层特征表达困难、中层特征拟合能力不足、深层特征依赖大数据样本且代表性特征挖掘不充分等关键技术，系统地构建了“多尺度同异质底层特征学习—主题特征深度拟合建模—深度主题特征融合的场景语义理解”方法体系，提出了基于全局上下文细节感知以及边缘监督的城市典型地物提取方法，实现了复杂城市背景下的城市典型地物的精确位置与复杂边缘提取；提出了基于多尺度稀疏先验约束与全局卷积长短期记忆模块的高光谱分类解混方法，实现了空谱联合解译的同时降低数据标注成本以及算法复杂程度；提出了基于代表性城市场景语义信息建模与迁移学习的城市土地利用制图框架，解决了基于深度学习的场景分类依赖大样本且代表性特征挖掘不充分等问题。.在项目的支持下，项目成员在IEEE TCYB，RSE，ISPRS P&RS，IEEE TGRS等国际学报上发表了14篇SCI学术论文，在IGARSS等国际会议上发表了16篇会议论文；该项目的研究成果成功应用于2022年武汉1+8城市圈违章用地分析，2022年河湖生态管控空间重点监管对象遥感智能识别技术研究，2022年绍兴市图像分割与特征提取的重要标识快速识别应用技术服务等；项目成员1人晋升副教授，培养研究生11名。