基于深度置信神经网络的地表覆盖质量检测方法研究

Land cover classification information is one of the important geographic information. The quality of land cover classification data will directly affect the decision making and the development of security services of geomatic information industry. The development of the quality automated inspection and testing technology in the field of land cover classification data will guarantee the efficiency and accuracy of decision-making services based on geographic information.Deep learning is the technology which builds machine learning models based on many hidden layers and vast amounts of training data to learn more useful features, and ultimately enhance the accuracy of the classification data or pattern recognition. The study of the project is to use the deep belief neural network for inspecting and testing the quality of land cover classification data of the high resolution remote sensing image. Firstly, based on the quality rules of inspection and testing of the image classification to extract and analysis of image feature, and based on the existing prior knowledge of the inspection and testing of the image classification to build the deep belief neural network. Secondly, input the large sample of characteristic data of the land cover image information to train the deep belief neural network and output the land cover classified image features to be tested, Finally, carry out the quality inspection and testing of land cover classified data using the trained deep relief neural network, and compare, verify and improve the algorithm of the project study based on the land cover classification data of national geographic census and monitoring.

地表覆盖分类信息作为一种重要的地理信息，其质量好坏将直接影响到测绘地理信息服务保障的质量及整个测绘地理信息行业的发展，研究一套高效、自动的地表覆盖分类质量检测的方法具有重大的意义。深度学习是通过构建具有很多隐层的机器学习模型和海量的训练数据，来学习更有用的特征，从而最终提升分类或模式识别的准确性。本项目利用深度置信神经网络对高分辨率遥感影像地表覆盖分类信息质量检测展开研究。通过研究基于影像的质量检测特征，提取与分析影像的特征；结合已有的先验知识，构建基于质量检测的深度置信神经网络；以大样本的遥感影像地表覆盖分类信息特征数据作为深度置信神经网络的输入，待检测类别信息作为输出，对网络进行训练，最终将训练好的深度置信神经网络用于地表覆盖分类信息的质量检测，并利用地理国情监测的地表覆盖分类数据对本研究的算法进行比较验证。

根据项目研究计划，本项目利用深度置信神经网络对高分辨率遥感影像地表覆盖分类信息质量检测展开研究，通过研究基于影像的质量检测特征，提取与分析影像的特征，结合已有的先验知识，构建基于质量检测的深度置信神经网络，开展基于深度置信神经网络的分类信息质量检测研究等。一定程度上解决在地表覆盖分类信息质量检测中存在的可靠性差、自动化程度低等问题，形成一套基于高分辨率遥感影像地表覆盖分类信息质量的自动检测方法。.在研究过程中项目组不断跟踪深度学习领域最新研究趋势，结合项目实际将网络类型转向深层卷积神经网络，以更好适应地表覆盖分类光学遥感影像的计算机视觉识别需求。.项目在模型训练中完成了适用于高分辨率地表覆盖分类信息质量检验特征向量的收集。顺利完成了适用于高分辨率地表覆盖分类信息质量检验的深度卷积神经网络模型的建立。提出了基于质量检测的深度卷积神经网络多源高分辨率遥感数据质量的自动缺陷性检测方法。.项目组按计划完成了全部研究任务，并整理研究成果发表SCI期刊论文6篇、中文核心期刊论文7篇、EI期刊论文2篇、EI会议论文5篇。培养硕士4名、博士4名。