Lisa Landuyt

Effective disaster response relies on rapid disaster response, where oblique aerial video is the primary modality for initial scouting due to its ability to maximize spatial coverage and situational awareness in limited flight time. However, the on-board processing of high-resolution oblique streams is severely bottlenecked by the strict Size, Weight, and Power (SWaP) constraints of Unmanned Aerial Vehicles (UAVs). The computational density required to process these wide-field-of-view streams precludes low-latency inference on standard edge hardware. To address this, we propose Temporal Token Reuse (TTR), an adaptive inference framework capable of accelerating video segmentation on embedded devices. TTR exploits the intrinsic spatiotemporal redundancy of aerial video by formulating image patches as tokens; it utilizes a lightweight similarity metric to dynamically identify static regions and propagate their precomputed deep features, thereby bypassing redundant backbone computations. We validate the framework on standard benchmarks and a newly curated Oblique Floodwater Dataset designed for hydrological monitoring. Experimental results on edge-grade hardware demonstrate that TTR achieves a 30% reduction in inference latency with negligible degradation in segmentation accuracy (< 0.5% mIoU). These findings confirm that TTR effectively shifts the operational Pareto frontier, enabling high-fidelity, real-time oblique video understanding for time-critical remote sensing missions

Oil spills in ports are particular cases of oil pollution in water environments that call for specific monitoring measures. Apart from the ecological threats that they pose, their proximity to human activities and the financial losses induced by disturbed port activities add to the need for immediate action. However, in ports, established methods based on short-wave infrared sensors might not be applicable due to the relatively low thickness of the oil layer, and satellite images suffer from insufficient spatial resolution, given the agglomeration of objects in ports. In this study, a lightweight ultraviolet (UV) camera was exploited in both controlled experiments and a real port environment to estimate the potential and limitations of UV imagery in detecting oil spills, in comparison to RGB images. Specifically, motivated by the scarce research literature on this topic, we set up experiments simulating oil spills with various oil types, different viewing angles, and under different weather conditions, such that the separability between oil and background (water) could be better understood and objectively assessed. The UV camera was also used to detect real-world oil spills in a port environment after installing it on a vessel for continuous monitoring. Various separability metrics between water and oil, computed in both scenarios (controlled experiments and port environment), show that the UV cameras have better potential than RGB in detecting oil spills in port environments.

The occurrence of litter in natural areas is nowadays one of the major environmental challenges. The uncontrolled dumping of solid waste in nature not only threatens wildlife on land and in water, but also constitutes a serious threat to human health. The detection and monitoring of areas affected by litter pollution is thus of utmost importance, as it allows for the cleaning of these areas and guides public authorities in defining mitigation measures. Among the methods used to spot littered areas, aerial surveillance stands out as a valuable alternative as it allows for the detection of relatively small such regions while covering a relatively large area in a short timeframe. In this study, remotely piloted aircraft systems equipped with multispectral cameras are deployed over littered areas with the ultimate goal of obtaining classification maps based on spectral characteristics. Our approach employs classification algorithms based on random forest approaches in order to distinguish between four classes of natural land cover types and five litter classes. The obtained results show that the detection of various litter types is feasible in the proposed scenario and the employed machine learning algorithms achieve accuracies superior to 85% for all classes in test data. The study further explores sources of errors, the effect of spatial resolution on the retrieved maps and the applicability of the designed algorithm to floating litter detection.

The economic loss associated with natural hazards has drastically increased over the past decades, reaching over $210 billion dollars worldwide in 2020. The explication of regional-scale climate change effects with the tendency to exacerbate local climate criticalities has long jeopardized disaster resilience and the coping capacity of many communities. There is a lack of a robust operational linkage between the pre-disaster and post-disaster segments when a disaster occurs. This hampers an effective emergency response often leading to delayed humanitarian intervention and unplanned evacuations. Moreover, the great amount of openly available impact information on past events is commonly discarded and the forecast potential which the data yields has yet to be fully explored. In this context, the Impact-based Forecasting (IbF) approach aims to interconnect pre-emptive planning for early action with post-disaster impacts while taking advantage of historical data. The underlying principle of IbF is that the magnitude of an event is translated to site-specific impact information. Therefore, a paradigm shift from the conventional magnitude-likelihood relationship to impact-likelihood is proposed. This research develops a method to fully exploit the potential of IbF while overcoming the typical site-specificity of emergency response through remote sensing and automation. While the IbF framework allows for a multi-hazard approach, here we present a method targeting the ex-ante impact assessment of riverine floods. The analysis consists of two main components: i) the delineation of the flood extent from Sentinel-1 SAR imagery and ii) the definition of the event impact on the population, land and built environment. The IbF impact-likelihood relationship is ultimately derived by matching the two components for a historical event series. A fully automated Google Earth Engine algorithm for flood extent mapping with a 10 m spatial resolution has been developed to detect floodwater with a single-scene classification based on an automated thresholding method. The flood magnitude is then matched with open-access geodata such as human settlements, population density, land cover and infrastructure from the OpenStreetMap catalogue to generate the impact assessment. Once trained on several site or region specific past events, it can automatically forecast the impact associated with a given event magnitude. Here we apply the technique to three case studies including the flooding associated with the Tropical Cyclone Idai, which made landfall in Mozambique in March 2019 causing over 1200 fatalities and $2 billion worth of damage. The performance of the flood mapping algorithm has been evaluated as satisfactory for the impact application and further validation at two additional sites is ongoing. Therefore, local triggers can be set to ensure a valuable temporal window to promptly plan and estimate the cost of intervention on the field. This work is a first step to providing a consistent and regionally transferable disaster preparedness tool that allows for multi-hazard impact forecasts.

As floods pose an increasing threat to our society, insights into their occurrence and dynamics are of major importance for emergency relief, damage assessment, the optimization of predictive models, and spatial planning. Due to their capability of providing synoptic observations independent of cloud cover and daylight, synthetic-aperture radar (SAR) sensors are an invaluable tool for flood mapping and monitoring. In this study, the potential of SAR, and more specifically Sentinel-1, for automated flood monitoring in Flanders is assessed. Its capability to detect floods with varying characteristics is investigated, and an approach for automated monitoring is presented. This approach, combining thresholding and region growing, requires a SAR image pair and several ancillary data layers, including elevation, land cover, and flood risk, as input. The resulting map discriminates permanent water, open flooding, long-term flooding, possible flooding, flooded vegetation, and possibly flooded forests from dry land. Invisible forested areas are indicated as well. A quantitative and qualitative accuracy assessment, based on 17 and 138 flood maps, respectively, highlights the approach's robustness and improved accuracy compared to benchmark techniques. Furthermore, main sources of confusion are identified and suggestions for future improvements are listed.

The European Space Agency’s Sentinel-1 constellation provides timely and freely available dual-polarized C-band Synthetic Aperture Radar (SAR) imagery. The launch of these and other SAR sensors has boosted the field of SAR-based flood mapping. However, flood mapping in vegetated areas remains a topic under investigation, as backscatter is the result of a complex mixture of backscattering mechanisms and strongly depends on the wave and vegetation characteristics. In this paper, we present an unsupervised object-based clustering framework capable of mapping flooding in the presence and absence of flooded vegetation based on freely and globally available data only. Based on a SAR image pair, the region of interest is segmented into objects, which are converted to a SAR-optical feature space and clustered using K-means. These clusters are then classified based on automatically determined thresholds, and the resulting classification is refined by means of several region growing post-processing steps. The final outcome discriminates between dry land, permanent water, open flooding, and flooded vegetation. Forested areas, which might hide flooding, are indicated as well. The framework is presented based on four case studies, of which two contain flooded vegetation. For the optimal parameter combination, three-class F1 scores between 0.76 and 0.91 are obtained depending on the case, and the pixel- and object-based thresholding benchmarks are outperformed. Furthermore, this framework allows an easy integration of additional data sources when these become available.

Insights into flood dynamics, rather than solely flood extent, are critical for effective flood disaster management, in particular in the context of emergency relief and damage assessment. Although flood dynamics provide insight in the spatio-temporal behaviour of a flood event, to date operational visualization tools are scarce or even non-existent. In this letter, we distil a flood dynamics map from a radar satellite image time series (SITS). For this, we have upscaled and refined an existing design that was originally developed on a small area, describing flood dynamics using an object-based approach and a graph-based representation. Two case studies are used to demonstrate the operational value of this method by visualizing flood dynamics which are not visible on regular flood extent maps. Delineated water bodies are grouped into graphs according to their spatial overlap on consecutive timesteps. Differences in area and backscatter are used to quantify the amount of variation, resulting in a global variation map and a temporal profile for each water body, visually describing the evolution of the backscatter and number of polygons that make up the water body. The process of upscaling led us to applying a different water delineation approach, a different way of ensuring the minimal mapping unit and an increased code efficiency. The framework delivers a new way of visualizing floods, which is straightforward and efficient. Produced global variation maps can be applied in a context of data assimilation and disaster impact management.

Flooding is one of the most severe natural hazards with respect to both economic and human losses. Therefore, it is of prime importance to provide timely and accurate information about these events, both for emergency management and flood risk assessment. Synthetic Aperture Radar (SAR) sensors are particularly suited to provide flood observations given their all weather, day/night sensing capability and the distinctive backscatter characteristics of smooth water surfaces. Over the past years, a considerable number of SAR-based flood mapping approaches has been developed. However, most of these focus on the retrieval of open water surfaces only. Flood mapping in vegetated and urban areas remains challenging due to the more complex backscatter mechanisms occurring in these areas. Yet, accurate delineation of floods in these areas is all the more important given their economical and societal relevance.This study focuses on the retrieval of flood extent information in complex, vegetated landscapes by means of freely available data. The considered imagery includes a pair of Sentinel-1 images, one acquired before and one acquired during the flood, as well as a cloud free Sentinel-2 image or mosaic acquired under non-flooded but representative vegetative conditions. An object-based change detection approach is used. Grouping pixels into segments prior to further analysis allows the integration of contextual and morphological information as well as the combination of different information sources. Segmentation is achieved by means of the quickshift algorithm, considering both polarization bands of the SAR image pair. Next, object properties with respect to SAR backscatter, surface reflectance and elevation are calculated and objects are grouped using spectral clustering. By including optical imagery, vegetation cover is considered and the flooded vegetation class can be better discriminated. The resulting clusters are then assessed, analysed and classified. Post-processing is done by means of contextual and topographical information. The use of fuzzy logic allows to assign an uncertainty measure to the obtained classification. For full scene processing, a thresholding-based preliminary flood extent is first derived in order to speed up the classification process and correct for class imbalance. The approach is presented based on multiple study cases, amongst which the 2019 floods along the Sava River, Croatia.