Simultaneous contextual classification of multitemporal and multiscale remote sensing imagery based on existing GIS data for training

该项目的目标是开发一种新的方法，在没有任何人工标记的训练数据的情况下，对多时间和多尺度遥感图像进行监督的基于上下文的分类。主要的科学贡献是开发了新的训练方法，这些方法可以容忍标签噪声，即相当数量的带有错误类标签的训练样本。使用这些方法，应该可以使用现有的土地覆盖（LC）数据来派生类标签，用于训练待分类图像的所有像素。我们建议使用自动生成的大量训练数据，以及可以处理这些数据中不可避免的错误的方法，而不是使用稀疏的手工标记的训练数据。该方法的数学框架由条件随机场（CRF）给出。考虑到LC数据在多个分辨率下将具有不同的类结构，我们将构建一个可以同时对多个时代和不同几何分辨率的数据进行分类的CRF。我们依靠全局、区域和局部LC数据集的存在来获得训练数据。我们将开发新的概率方法来考虑训练中的标签噪声，以便不仅获得连接CRF的未知类标签与数据的分类器的参数，而且还获得连接不同时代图像的参数。作为一个重要贡献，我们将考虑LC数据中的误差是空间相关的这一事实。建议的项目构成了基于图的图像分类背景下标签噪声容忍训练程序原则的第一个应用，也是考虑在不同语义细节级别建模的对象之间相互作用的最通用技术之一。因此，应该有可能降低全球或区域和本地LC数据集更新的成本，例如，通过使用低分辨率的廉价图像来获得高分辨率数据变化的提示。新方法在实际数据上进行了评估，并参考了人工生成的参考数据。在与中国国家信息中心（NGCC）现有的谅解备忘录的框架下，我们将在德国和中国的不同测试地点研究该方法。我们将使用NGCC免费开发的30 m几何分辨率的全球土地覆盖数据集GLC30作为我们测试用例的粗分辨率数据集。我们将使用的高分辨率数据集分别来自德国测量局和NGCC。

It is the goal of the proposed project to develop a novel methodology for the supervised context-based classification of multitemporal and multiscale remote sensing imagery without any manually labelled training data. The main scientific contribution is the development of new training methods that are tolerant to label noise, i.e., to a considerable amount of training samples with erroneous class labels. Using these methods it should become possible to use existing land cover (LC) data to derive class labels to be used for training for all pixels of an image to be classified. Rather than using sparse hand-labelled training data, we propose using an abundance of training data generated automatically, along with methods that can deal with the inevitable errors in these data. The mathematical framework for the proposed methodology is given by Conditional Random Fields (CRF). We will build a CRF that can classify data from multiple epochs and having different geometrical resolutions simultaneously, considering the fact that LC data at multiple resolutions will be characterised by different class structures. We rely on the existence of both, global, regional and local LC data sets to derive training data. We will develop new probabilistic approaches for considering label noise in training in order to obtain not only the parameters of the classifiers linking the unknown class labels of the CRF with the data, but also the parameters linking the images at different epochs with each other. As an important contribution we will consider the fact that errors in LC data are spatially correlated. The suggested project constitutes the first application of the principles of label-noise tolerant training procedures in the context of graph-based image classification, and one of the most general techniques for considering interactions between objects modelled at different semantic levels of detail. As a consequence, it should become possible to cut the costs for the update of global or regional and local LC data sets, e.g. by using cheap imagery of low resolution to get hints for changes in the high-resolution data. The new methodology is evaluated on real data with a reference that was generated manually. In the frame of an existing Memorandum of Understanding with the National Geomatics Center of China (NGCC) we will investigate the methodology in different test sites in Germany and China. We will use the global land cover data set GLC30 with 30 m geometrical resolution, developed by NGCC and available free of charge, as the coarse-resolution data set in our test cases. The high-resolution data sets we will use are those from the German Survey Authorities and NGCC, respectively.