Due to its all time capability, synthetic aperture radar (SAR) remote sensing plays an important role in Earth observation. The ability to interpret the data is limited, even for experts, as the human eye is not familiar to the impact of distance-dependent imaging, signal intensities detected in the radar spectrum as well as image characteristics related to speckle or steps of post-processing. This paper is concerned with machine learning for SAR-to-optical image-to-image translation in order to support the interpretation and analysis of original data. A conditional adversarial network is adopted and optimized in order to generate alternative SAR image representations based on the combination of SAR images (starting point) and optical images (reference) for training. Following this strategy, the focus is set on the value of empirical knowledge for initialization, the impact of results on follow-up applications, and the discussion of opportunities/drawbacks related to this application of deep learning. Case study results are shown for high resolution (SAR: TerraSAR-X, optical: ALOS PRISM) and low resolution (Sentinel-1 and -2) data. The properties of the alternative image representation are evaluated based on feedback from experts in SAR remote sensing and the impact on road extraction as an example for follow-up applications. The results provide the basis to explain fundamental limitations affecting the SAR-to-optical image translation idea but also indicate benefits from alternative SAR image representations.

Studies in the last years have proved the outstanding performance of deep learning for computer vision tasks in the remote sensing field, such as disparity estimation. However, available datasets mostly focus on close-range applications like autonomous driving or robot manipulation. To reduce the domain gap while training we present SyntCities, a synthetic dataset resembling the aerial imagery on urban areas. The pipeline used to render the images is based on 3D modelling, which helps to avoid acquisition costs, provides sub-pixel accurate dense ground truth and simulates different illumination conditions. The dataset additionally provides multi-class semantic maps and can be used converted to point cloud format to benefit a wider research community. We focus on the task of disparity estimation and evaluate the performance of the traditional Semi-Global Matching and state of the art architectures, trained with SyntCities and other datasets, on real aerial and satellite images. A comparison with the widely used SceneFlow dataset is also presented. Strategies using a mixture of both real and synthetic samples are studied as well. Results show significant improvements in terms of accuracy for the disparity maps. SyntCities can be downloaded at: https://tinyurl.com/77e3n6m9

Deep learning is a powerful tool to extract both individual building and roof plane polygons. But deep learning requires a large amount of labeled data. Hence, publicly available level of detail (LoD)-2 datasets are a natural choice to train fully convolutional neural networks (FCNs) models for both building section and roof plane instance segmentation. Since publicly available datasets are often automatically derived, e.g. based on laser scanning, they lack on annotation accuracy. To complement such a dataset, we introduce manually annotated and synthetically generated data. Manually annotated data is domain-specific and has a high annotation quality but is expensive to obtain. Synthetically generated data has high-quality annotations by definition, but lacks domain-specificity. Moreover, we not only detect individual building section instances, but also roof plane instances. We predict separations not only between individual buildings, but also by a class that describes the line which separates roof planes. The predicted building and roof plane instances are polygonized by a simple tree search algorithm. To achieve more regular polygons, we utilize the Douglas-Peucker polygon simplification algorithm. We describe our dataset in detail to allow comparability between successive methods. To facilitate future works in building and roof plane prediction, our Roof3D dataset is accessible at https://github.com/dlrPHS/GPUB.

The creation of digital surface models (DSMs) from aerial and satellite imagery is often the starting point for different remote sensing applications. For this task, the two main used approaches are stereo matching and multi-view stereo (MVS). The former needs stereo-rectified pairs as inputs and the results are in the disparity domain. The latter works with images from various perspectives and produces a result in the depth domain. So far, both approaches have proven to be successful in producing accurate DSMs, especially in the deep learning area. Nonetheless, an assessment between the two is difficult due to the differences in the input data, the domain where the directly generated results are provided and the evaluation metrics. In this manuscript, we processed synthetic and real optical data to be compatible with the stereo and MVS algorithms. Such data is then applied to learning-based algorithms in both analyzed solutions. We focus on an experimental setting trying to establish a comparison between the algorithms as fair as possible. In particular, we looked at urban areas with high object densities and sharp boundaries, which pose challenges such as occlusions and depth discontinuities. Results show in general a good performance for all experiments, with specific differences in the reconstructed objects. We describe qualitatively and quantitatively the performance of the compared cases. Moreover, we consider an additional case to fuse the results into a DSM utilizing confidence estimation, showing a further improvement and opening up a possibility for further research.

Dense matching plays a crucial role in computer vision and remote sensing, to rapidly provide stereo products using inexpensive hardware. Along with the development of deep learning, the Guided Aggregation Network (GA-Net) achieves state-of-the-art performance via the proposed Semi-Global Guided Aggregation layers and reduces the use of costly 3D convolutional layers. To solve the problem of GA-Net requiring large GPU memory consumption, we design a pyramid architecture to modify the model. Starting from a downsampled stereo input, the disparity is estimated and continuously refined through the pyramid levels. Thus, the disparity search is only applied for a small size of stereo pair and then confined within a short residual range for minor correction, leading to highly reduced memory usage and runtime. Tests on close-range, aerial, and satellite data demonstrate that the proposed algorithm achieves significantly higher efficiency (around eight times faster consuming only 20–40% GPU memory) and comparable results with GA-Net on remote sensing data. Thanks to this coarse-to-fine estimation, we successfully process remote sensing datasets with very large disparity ranges, which could not be processed with GA-Net due to GPU memory limitations.