Remote sensing is extensively used in cartography. As transportation networks grow and change, extracting roads automatically from satellite images is crucial to keep maps up-to-date. Synthetic aperture radar (SAR) satellites can provide high-resolution topographical maps. However, roads are difficult to identify in these data as they look visually similar to targets, such as rivers and railways. Most road extraction methods on SAR images still rely on a prior segmentation performed by the classical computer vision algorithms. Few works study the potential of deep learning techniques, despite their successful applications to optical imagery. This letter presents an evaluation of fully convolutional neural networks (FCNNs) for road segmentation in SAR images. We study the relative performance of early and state-of-the-art networks after carefully enhancing their sensitivity toward thin objects by adding the spatial tolerance rules. Our models show promising results, successfully extracting most of the roads in our test data set. This shows that although FCNNs natively lack efficiency for road segmentation, they are capable of good results if properly tuned. As the segmentation quality does not scale well with the increasing depth of the networks, the design of specialized architectures for roads extraction should yield better performances.
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.
29 March 2018 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Tasks such as the monitoring of natural disasters or the detection of change highly benefit from complementary information about an area or a specific object of interest. The required information is provided by fusing high accurate coregistered and georeferenced datasets. Aligned high-resolution optical and synthetic aperture radar (SAR) data additionally enable an absolute geolocation accuracy improvement of the optical images by extracting accurate and reliable ground control points (GCPs) from the SAR images. In this paper, we investigate the applicability of a deep learning based matching concept for the generation of precise and accurate GCPs from SAR satellite images by matching optical and SAR images. To this end, conditional generative adversarial networks (cGANs) are trained to generate SAR-like image patches from optical images. For training and testing, optical and SAR image patches are extracted from TerraSAR-X and PRISM image pairs covering greater urban areas spread over Europe. The artificially generated patches are then used to improve the conditions for three known matching approaches based on normalized cross-correlation (NCC), scale-invariant feature transform (SIFT), and binary robust invariant scalable key (BRISK), which are normally not usable for the matching of optical and SAR images. The results validate that a NCC-, SIFT-, and BRISK-based matching greatly benefit, in terms of matching accuracy and precision, from the use of the artificial templates. The comparison with two state-of-the-art optical and SAR matching approaches shows the potential of the proposed method but also revealed some challenges and the necessity for further developments.
Improving the geo-localization of optical satellite images is an important pre-processing step for many remote sensing tasks like monitoring by image time series or scene analysis after sudden events. These tasks require geo-referenced and precisely co-registered multi-sensor data. Images captured by the high resolution synthetic aperture radar (SAR) satellite TerraSAR-X exhibit an absolute geo-location accuracy within a few decimeters. These images represent therefore a reliable source to improve the geo-location accuracy of optical images, which is in the order of tens of meters. In this paper, a deep learning-based approach for the geo-localization accuracy improvement of optical satellite images through SAR reference data is investigated. Image registration between SAR and optical images requires few, but accurate and reliable matching points. These are derived from a Siamese neural network. The network is trained using TerraSAR-X and PRISM image pairs covering greater urban areas spread over Europe, in order to learn the two-dimensional spatial shifts between optical and SAR image patches. Results confirm that accurate and reliable matching points can be generated with higher matching accuracy and precision with respect to state-of-the-art approaches.
5 December 2023 in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
This paper focuses on person detection in aerial and drone imagery, which is crucial for various operations such as situational awareness, search and rescue, and safe delivery of supplies. We aim to improve disaster response efforts by enhancing the speed, safety, and effectiveness of the process. Therefore, we introduce a new person detection dataset comprising 311 annotated aerial and drone images, acquired from helicopters and drones in different scenes, including urban and rural areas, and for different scenarios, such as estimation of damage in disaster-affected zones, and search and rescue operations in different countries. The amount of data considered and level of detail of the annotations resulted in a total of 10,050 annotated persons. To detect people in aerial and drone images, we propose a multi-stage training procedure to improve YOLOv3’s ability. The proposed procedure aims at addressing challenges such as variations in scenes, scenarios, people poses, as well as image scales and viewing angles. To evaluate the effectiveness of our proposed training procedure, we split our dataset into a training and a test set. The latter includes images acquired during real search and rescue exercises and operations, and is therefore representative for the challenges encountered during operational missions and suitable for an accurate assessment of the proposed models. Experimental results demonstrate the effectiveness of our proposed training procedure, as the model’s average precision exhibits a relevant increase with respect to the baseline value.
29 December 2022 in IET Intelligent Transport Systems
Parking is a vital component of today's transportation system and descriptive data are therefore of great importance for urban planning and traffic management. However, data quality is often low: managed parking places may only be partially inventoried, or parking at the curbside and on private ground may be missing. This paper presents a processing chain in which remote sensing data and statistical methods are combined to provide parking area estimates. First, parking spaces and other traffic areas are detected from aerial imagery using a convolutional neural network. Individual image segmentations are fused to increase completeness. Next, a Gamma hurdle model is estimated using the detected parking areas and OpenStreetMap and land use data to predict the parking area adjacent to streets. A systematic relationship is found between the road length and type and the parking area obtained. It is suggested that these results are informative to those needing information on parking in structurally similar regions.