Building in permafrost environments is challenging. For example, the bearing capacity of some permafrost soils may significantly decrease with warming and thawing due to increased unfrozen water content. Moreover, the seasonal freeze/thaw cycle of the active layer (AL) may lead to significant frost action and subsequently potential damage of infrastructure foundations.
Often, lack of equipment, time and financial resources hinder thorough surveys of the soil conditions (such as frost susceptibility), which should occur in the planning and design phase of construction projects. Additionally, permafrost degradation results in Active Layer thickness increase. This poses problems to currently stable situations where the foundations still lie within the permanently frozen and bearing-capable soils.
In the AALM4INFRAM ESA project satellite based remote sensing (Sentinel-1 InSAR and Sentinel-2 surface classification) was used together with geotechnical field surveys and geographic information systems (GIS) to
i. map the current amplitude of seasonal frost/thaw cycle induced ground movements,
ii. map frost susceptibility,
iii. estimate the permafrost degradation rate

As outcome of the project, maps showing the seasonal frost heave and the multi-annual permafrost degradation trend as a non-reversible motion component were produced for test sites at Ilulissat, Sisimiut, Kangerlussuaq and Qaanaaq. This gives a geographic overview of the soil parameters and the distribution of potentially difficult foundation soils in settlements in Greenland. Additionally, a combination of an AL-thickness estimation with a map of a measure of frost-susceptibility could be derived for Ilulissat. Data is distributed amongst important Greenlandic stakeholders on governmental, municipal and private level, from a web map platform that is also open for the public.
In general, the stakeholder-feedback was positive. Several critical areas with known permafrost-related damages were correctly mapped. Furthermore, unknown areas were identified and will be validated in the course of future site investigations for infrastructure projects.
In our contribution, we will present an overview on the methods used to derive the above stated map products from remote sensing perspective, discuss the resulting products in light of the current infrastructure damage problematic and give an outlook on future perspective of the presented approach in arctic infrastructure management.

According to the announcement of German Federal Ministry of Transport and Digital Infrastructure, Building Information Modeling (BIM) has become the standard digital format for the construction project of high-building and infrastructure in Germany after 2020. To contribute to the building monitoring in future, we are granted in an innovative project BIMSAR from German Federal Ministry for Economic Affairs and Energy. Our aim is to fuse the BIM- and SAR-data to realize a semantic building monitoring at a fine scale, e.g., for change detection or deformation analysis.

Here BIM will be the reference basis for a digital building model. Via Persistent Scatterer Interferometry (PSI), PS-points are extracted from high-resolution SAR image stack. These PS-points are regarded as pseudo-substructures, of which a building consists. Their important properties contain geographic position and deformation estimates. The PS-points will then be fused with the BIM-models and clustered to different self-contained units like façade and roof. This step will be implemented via a novel distance metric adapted by a dimension reduction, if needed, as well as AI-based algorithms. The fused data will be finally integrated into a web- and cloud-based monitoring platform. Except visualization, this platform also contains local and expert knowledge to assist users in data analysis and decision making.

In this presentation, we will demonstrate our test products to the fellows and share our ideas for semantic building monitoring. The first test object is a building complex – Olympiastützpunkt in Bochum, Germany. We have produced a BIM-model compatible with Industry Foundation Classes (IFC) format (International Standard ISO 16739-1:2018). The PS-points have been extracted from 29 TerraSAR-X images (2 m × 2 m). An initial platform was framed and is being tested. We will fuse and integrate the data into this platform. Meanwhile, we are tasking TerraSAR-X to collect very-high-resolution data (resolution of 0.25 m) for our second test building complex – Zeche Westfalen Ahlen. Last but not least, we will for the first time display a BIM-City and discuss the potential applications. Our future work is to improve our products based on user feedback and to release a service ready for marketing.

Increasing demand for green technologies is driving the need for more minerals and metals. If we are to
avoid the existential threats posed by climate change and ecological collapse, and remain within the Paris
Agreements of global average temperatures of well below 2◦C, an estimated 3 billion tons of metals will be
required to enable our energy [1]. The often forgotten consequence of such increased mining activities is the
question of what do we do with the resulting mine waste? Usually, mine waste are stored in dams called
‘tailings dams’. There is an urgent need to protect the surrounding communities, habitats, ecosystems from
another environmental disasters such as the recent Brumadinho failure of 2019, Brazil [2].
From hundreds of km in space, satellite-Interferometric Synthetic-Aperture Radar (InSAR) analysis allow
detection of millimeter scale ground motion. In this paper, ground deformation measurements from InSAR
and ground-based prism monitoring are compared to Finite Element (FE) simulation results for the Cadia
tailings dam collapse of 2018 in Australia. A coupled-consolidation analysis was used to predict the behaviour
of the tailings dam during various stages of the construction. The FE modelling results were in broad
agreement with the other deformation measurements, both in terms of magnitude and trends. The results,
however, deviate in the time immediately preceding the failure, following construction of the dam buttresses.
Moreover, the FE modelling is sensitive to parameter uncertainties such as spatial variability of the foundation soil. The 2D FE model is in better agreement in some parts of the failure area than others. We are able
to detect anomalous deformation behaviour from InSAR within the slump area compared to the other areas
of the dam, potentially months before failure. Finally, we have demonstrated the complex spatio-temporal
variability of the dam deformation that is not captured in traditional, point-based monitoring approaches.
This study illustrates the complementarity of geotechnical and remote sensing techniques for the monitoring
of tailings dams.

References

[1] K Hund, D La Porta, TP Fabregas, T Laing, and J Dexhage. Minerals for climate action: the mineral
intensity of the clean energy transition. climate-smart mining facility, 2020.
[2] P. Robertson, L. de Melo, D. J. Williams, and G. W. Wilson. Report of the expert panel on the technical
causes of the failure of feij˜ao dam i. http://www. b1technicalinvestigation.com/, 2020.
1

Counting small objects like cars in high resolution satellite imagery using machine learning techniques is still a challenging task. A typical car can be represented in the 50cm resolution imagery as less than 10 pixels therefore the standard object detection techniques are not very effective. The difficulty is compounded by the lack of pre-annotated datasets composed of cars seen from overhead and representing a satellite grade resolution.
One possible solution to the above problem is to build and train a model that would detect cars in higher than satellite resolution (aerial - 10-15cm resolution) and then reuse the model and train it on synthetic dataset constructed from high resolution images.
As part of our project, we were interested in counting the new cars stored in cargo port terminal on densely populated parking. We acquired high resolution aerial imagery for a number of cargo ports around the world and annotated the training dataset. The human-annotated data constituted the ground truth. The labelled images were then converted into a density map.
For the training a deep multi-column convolutional neural network was used to learn the target density maps. The inputs are 256*256 images. We use Adam optimiser with a learning rate of 0.0001. Relu activation is use at all layers, except the final where linear activation is used. At this stage not regularisation is used, however in future we intend to investigate it utility. All layers were initialised with Xavier initialiser. The mean square error (MSE) was used as the loss function. Training was done for a total of 1000 epochs.
A small part of the annotated dataset was reserved for validation purposes and not included in the training dataset. Throughout this project, the R-value and R2-value were used to evaluate the accuracy of the trained model on the test data.
As a result, correlation coefficient of 98% and the R2 value of was 96% achieved and the first step of creating high-accuracy model working with aerial grade imagery was successfully completed. The highly accurate model was then used to generate the ground truth for not annotated satellite images and generate additional 8500+ high-resolution image titles.

As a next step the aerial images were downscaled to the resolution of interest (0.5 m/pixel) and the synthetic dataset was created. The Gaussian filter was used to blur images and the Bicubic interpolation was applied.
The synthetic dataset was then used to train a deep-learning model. The model takes a satellite grade resolution image as an input and returns the corresponding density map.
Transfer learning was used to leverage on the model that was already created for high (aerial) resolution images. The identical network was used but it was initialized with the weights of the high resolution trained model. The model was trained for 1400 epoch and a batch size of 32 was used.
The new model trained on synthetic low resolution images achieved in tests R2=0.95 and R=0.91. The good results can be attributed to the increase in both volume and the variety of training dataset.
The next step was to test the satellite grade resolution model on the actual satellite images. The fact that model is based on synthetic images meant that the training dataset images differed significantly from the actual satellite images.
In order to reduce the difference, the same treatment that was applied to the synthetic imagery was applied to the actual images. The gaussian filter was applied to the actual satellite image to down sample it by the factor of 2 and then the image was upscaled back to its original resolution.
The model trained on the synthetic low-resolution images was then used to make inference on the processed satellite imagery.
As the imagery used for testing was difficult to annotate as cars were barely visible to the human, a super resolution model was applied (Very Deep Super-Resolution for Geospatial Data) to create a high-resolution version of the images. The comparison of the human count to the counts generated by the model showed the accuracy of 99% was achieved.
The very high accuracy achieved shows the promise in application of transfer learning techniques for object detection, object counting and density estimation based on satellite imagery. The above model was tested on Planet's SkySat imagery both before and after lowering the orbit.
Skytek uses the above-described model to provide insights into risk accumulation in cargo ports for insurance industry.

With the ESA project SUMO4Rail an innovative user driven approach for optimising the monitoring capabilities and innovating the inspection process in terms of ground deformations over the German railway network was realised. This comprised an analysis of satellite based, interferometric measurements of existing deformation phenomena plus a predictive assessment of deformation risks at or in the vicinity of rail tracks.

In the framework of the project user requirements and needs of railway operators and other stakeholders were specified by the German Centre for Rail Traffic Research at the Federal Railway Authority (EBA) in cooperation with the Institute of Shipping Economics and Logistics (ISL). Subsequently existing ground motion data, produced for the German Ground Motion Service (BBD) and published by the Federal Institute for Geosciences and Natural Resources in Germany (BGR) were intensively analysed. The evaluation of the deformation information, respectively the identification of dynamic phenomena therein, was performed considering the three requirement based use cases. Each identified use case has certain requirements for the technical analysis, such as spatial and temporal resolution:

• Enhancement of infrastructure planning (“Planning assistant”)
• Monitoring of pre-known subsidence (“Maintenance assistant”)
• Identification of areas of potential subsidence risks (“Early-warning system”)

For the analysis exercise an area of interest (AOI) comprising the Duisburg Harbour and Hinterland with its complex and wide ranging rail network and logistics chain were specified. The AOI is fully covered by the BBD service providing deformation or displacement time series dating back to 2014. This service utilises the complete Sentinel-1 archive over Germany which is currently and continuously processed using PSI-WAP (Persistent Scatterer Inferferometry-Wide Area Product) technology, resulting in deformation products with 6 day temporal sampling both in ascending and descending acquisition geometry. For SUMO4Rail a time series of BBD information ranging from 2014 to 2018 was analysed.

Addressing the application of all developed processes in the project to the German railway network, robust prototypes and procedures were conceptualised and implemented to guarantee a high degree of automation for analysing the extensive deformation data basis. Analysis itself focusses on the generation of Active Deformation Areas (ADA) and decomposed terrain motion (East-West and Up-Down components) derived from different image geometries plus a deformation classification of time series integrating ancillary information like digital terrain models, slope, geology, hydrology and railway characteristica. Main findings show consistency and high correlation of deformation phenomena derived from different geometries and time series. The outcome of all analysis results and related developments will be discussed in this paper.

All results and process or system developments of SUMO4Rail were validated by the German rail operator DB Netz AG and are in general considered suitable to provide enhanced monitoring of railway infrastructure to detect damages at an early stage, thus enabling effective countermeasures, and to provide predictive detection of deformation risks or subsidence damages at or in the vicinity of rail tracks.
Deficits can be seen in the spatial and, in some cases, temporal resolution, as the requirements along railway lines are very specific due to the special characteristics (linear and narrow shape) of railway tracks. These challenges could be addressed by combining BBD data with additional high resolution data. Technical and practical shortcomings were identified in the homogenization of multiple PSI datasets and Vertical-and Horizontal Displacement (E-W) point maps, specifically in the domain of displacement time series which needs to be addressed on a technical or engineering level. Furthermore data handling of extensive data volumes, as exemplarily exercised in SUMO4Rail, requires efficient hardware infrastructures and processing capacities, also emphasising the dissemination of information to the user.

SNCF Réseau is the owner and main manager of the French national railway network. SNCF Réseau operates 30,000 km of lines in service throughout all of France with about 94% of conventional lines and 6% of high-speed lines (HSL), both monitored and maintained 24/7.

The current earthwork monitoring survey is performed using different methods: foot patrols for visual inspection, track geometry monitoring (via a dedicated measurement train), topographic survey, local investigation such as inclinometric or geophysical surveys and high-resolution InSAR studies on specific sites. However, SNCF Réseau does not currently use InSAR technique to have an overview of ground motions across its network. Therefore, SNCF Réseau was highly interested in taking part in the project led by ESA « EO4Infrastructure (Infrastructure Mapping and Planning) » as an end-user. TRE ALTAMIRA will generate ground motion maps over a specific area selected by SNCF Réseau in the East of France. This area is 300-km-long and 100-km-wide, spanning from Reims to Strasbourg. It includes railways of different categories (i.e., high-speed and classic lines). This area is affected by motions of different types (e.g., subsidence, landslides), which are already under in-situ monitoring (via topographic survey for instance).

The processing will be based on the patented SqueeSAR® algorithm using Sentinel-1 Medium-Resolution images for a regional-scale study, and COSMO-SkyMED High-Resolution images for a city-scale study around Reims. This will allow an assessment of high-resolution imagery contribution to infrastructure monitoring.

In addition, the project will allow to understand the impact of land-use changes on floods and landslides using Sentinel-2 images and OSO maps, with seasonal laps of time, combined with NIR (Near InfraRed) images data for soil moisture content (a main component in landslide predictability).

The aim of the validation phase is to compare InSAR studies and ground motion maps with in-situ data. The overall goal is to have a critical analysis and therefore to assess the relevance, utility, and value of the usability of this method as a survey tool over the earthworks of the railway network.