Publications in the field of Earth observation to support humanitarian operations
2024
International policy and humanitarian guidance emphasize the need for precise, subnational malaria risk assessments with cross-regional comparability. Spatially explicit indicator-based assessments can support humanitarian aid organizations in identifying and localizing vulnerable populations for scaling resources and prioritizing aid delivery. However, the reliability of these assessments is often uncertain due to data quality issues. This article introduces a data evaluation framework to assist risk modelers in evaluating data adequacy. We operationalize the concept of “data adequacy” by considering “quality by design” (suitability) and “quality of conformance” (reliability). Based on a use case we developed in collaboration with Médecins Sans Frontières, we assessed data sources popular in spatial malaria risk assessments and related domains, including data from the Malaria Atlas Project, a healthcare facility database, WorldPop population counts, Climate Hazards group Infrared Precipitation with Stations (CHIRPS) precipitation estimates, European Centre for Medium-Range Weather Forecasts (ECMWF) precipitation forecast, and Armed Conflict Location and Event Data Project (ACLED) conflict events data. Our findings indicate that data availability is generally not a bottleneck, and data producers effectively communicate contextual information pertaining to sources, methodology, limitations and uncertainties. However, determining such data’s adequacy definitively for supporting humanitarian intervention planning remains challenging due to potential inaccuracies, incompleteness or outdatedness that are difficult to quantify. Nevertheless, the data hold value for awareness raising, advocacy and recognizing trends and patterns valuable for humanitarian contexts. We contribute a domain-agnostic, systematic approach to geodata adequacy evaluation, with the aim of enhancing geospatial risk assessments, facilitating evidence-based decisions.
Updated building footprints with refugee camps from high-resolution satellite imagery can support related humanitarian operations. This study explores the utilization of the “Segment Anything Model” (SAM) and one of its branches, SAM-Adapter, for semantic segmentation tasks in the building extraction from satellite imagery. SAM-Adapter is a lightweight adaptation of the SAM and emerges as a powerful tool for this extraction task across diverse refugee camps. Our research proves that SAM-Adapter excels in scenarios where data availability is limited compared to other classic (e.g., U-Net) or advanced semantic segmentation models (e.g., Transformer). Furthermore, the impact of upscaling techniques on model performance is highlighted, with methods like super-resolution (SR) models proving invaluable for improving model performance. Additionally, the study unveils intriguing phenomena, including the model’s rapid convergence in the first training epoch when using upscaled image data for training, suggesting opportunities for future research. The codes covering each step from data preparation, model training, model inferencing, and the generation of Shapefiles for predicted masks are available on a GitHub repository to benefit the extended scientific community and humanitarian operations.
2023
CustomizingSegment Anything Model(SAM)hasrecentlyat-tractedconsiderableattention in remote sensing domains.This studyex-plores the performance of SAM-Adapter in refugee-dwelling extractioninthree different refugee camps from high-resolution satelliteimages.Thefindings indicate that with scarce sample data, SAM-Adaptermarginallyoutperforms other semantic segmentation models. This underscores SAM’spromising potential forbuilding extraction tasks when data is limited.
Even though computer vision models are excellent for automatic scene segmentation and object identification from remotely sensed imagery, they demand a huge corpus of annotated data for the training and validation which is a huge challenge in humanitarian emergency response. To tackle this problem, we propose unsupervised dwelling object counting combining Variational Autoencoder (VAE) with an anomaly detection approach. The approach is tested in six Forcibly Displaced People (FDP) settlement areas situated in different parts of the world. Using an anomaly map computed with the VAE model, we demonstrated the possibility of properly locating dwelling objects using anomaly maps. Dwelling counts are obtained by further segmenting anomaly maps. Results show that, though it has strong spatio-temporal variation, the VAE model exhibits promising potential for locating and counting dwellings. It is also observed that in FDP settlements with dense buildings and extremely low contrast between buildings and ground or environment, the performance is relatively lower than the performance achieved in settlement areas with regularly spaced and less complex building structures.
Even though computer vision models are excellent for automatic scene segmentation and object identification from remotely sensed imagery, they demand a huge corpus of annotated data for the training and validation which is a huge challenge in humanitarian emergency response. To tackle this problem, we propose unsupervised dwelling object counting combining Variational Autoencoder (VAE) with an anomaly detection approach. The approach is tested in six Forcibly Displaced People (FDP) settlement areas situated in different parts of the world. Using an anomaly map computed with the VAE model, we demonstrated the possibility of properly locating dwelling objects using anomaly maps. Dwelling counts are obtained by further segmenting anomaly maps. Results show that, though it has strong spatio-temporal variation, the VAE model exhibits promising potential for locating and counting dwellings. It is also observed that in FDP settlements with dense buildings and extremely low contrast between buildings and ground or environment, the performance is relatively lower than the performance achieved in settlement areas with regularly spaced and less complex building structures.
As the climate crisis continues to worsen, there is an increasing demand for scientific evidence from Climate Risk and Vulnerability Assessments (CRVA). We present 12 methodological advancements to the Impact Chain-based CRVA (IC-based CRVA) framework, which combines participatory and data-driven approaches to identify and measure climate risks in complex socio-ecological systems. The advancements improve the framework along five axes, including the existing workflow, stakeholder engagement, uncertainty management, socio-economic scenario modeling, and transboundary climate risk examination. Eleven case studies were conducted and evaluated to produce these advancements. Our paper addresses two key research questions: a) How can the IC-based CRVA framework be methodologically advanced to produce more accurate and insightful results? and b) How effectively can the framework be applied in research and policy domains that it was not initially designed for? We propose methodological advancements to capture dynamics between risk factors, to resolve contradictory worldviews, and to maintain consistency between Impact Chains across policy scales. We suggest using scenario-planning techniques and integrating uncertainties via Probability Density Functions and Reverse Geometric Aggregation. Our research examines the applicability of IC-based CRVAs to address transboundary climate risks and integrating macro-economic models to reflect possible future socio-economic exposure. Our findings demonstrate that the modular structure of IC-based CRVA allows for the integration of various methodological advancements, and further advancements are possible to better assess complex climate risks and improve adaptation decision-making.
Dwelling information is very important for various applications in humanitarian emergency response. For this, Earth observation is crucial to have spatially explicit and temporally frequent observations. Coupled with advances in computer vision, especially with the proliferation of state-of-the-art deep learning models, are providing a new opportunity for automatic information retrieval from remotely sensed imagery. Despite their proven performance, they have two known limitations, viz, the requirement of intensive data for training and lack of universal generalization under changing scene characteristics and respective data distributions. To tackle this problem, the current study has investigated the relevance of a meta-learning approach for the creation of a spatially transferable optimal model for dwelling extraction in IDP/refugee settlements. The study followed a Model Agnostic meta-Learning (MAML) strategy with newly designed and tested variates with weighted loss gradient update plus self-supervision in the adaptation phase to the target locations using a few samples. The approach is tested using multi-sensor, multi-temporal satellite imagery from eight IDP settlements. Furthermore, a thorough investigation is undertaken on task-specific transfers and their association with deep-embedded feature space and image structural similarity. Results indicate that for some target sites, task-specific transfers perform better than MAML approaches. When MAML is trained with a weighted loss gradient update, it yielded better performance. The best performance (MIoU 0.623 and an F-1 score of 76.7 %) was achieved when MAML is aided with self-supervision using pseudo-labels from unlabelled target data. In all experimental setups, though increasing adaptation samples contribute to positive transfer, the marginal contribution from additional samples is decreasing and stagnates when the adaptation sample size reaches ∼ 35 % of the target dataset.
2022
The increasing number of satellites offering daily image acquisition coverage at very high resolution (VHR) provides new insights into the dynamics of fine-scale changes at temporary settlements. This article presents first experiences with spaceborne mapping of refugee settlements using daily revisiting radar (ICEYE) and optical (PlanetScope) imagery of a study area in the Northeastern Democratic Republic of Congo, which hosts villages and informal dwellings of forcibly displaced persons. The information content of the sensors is systematically compared with respect to mapping land use, delineating built-up areas, and identifying camps. Validation against independent reference data shows many of the lightweight dwellings to not be included in the radar image due to the use of natural construction materials. The best distinctions between built-up areas and bare soil was achieved upon merging optical and radar data, resulting in a classification accuracy of 94.1%. The classification is based on four predefined classes, and more accurately identifies urban structures compared to the two-class approach that either underestimates or overestimates built-up areas. Accuracies between 83.4 % and 92.8 % have been achieved for 10 identified campsites. The results are a promising step toward operationally identifying refugee settlements for emergency response.
This research assessed the influences of four band combinations and three types of pretrained weights on the performance of semantic segmentation in extracting refugee dwelling footprints of the Kule refugee camp in Ethiopia during a dry season and a wet season from very high spatial resolution imagery. We chose a classical network, U-Net with VGG16 as a backbone, for all segmentation experiments. The selected band combinations include 1) RGBN (Red, Green, Blue, and Near Infrared), 2) RGB, 3) RGN, and 4) RNB. The three types of pretrained weights are 1) randomly initialized weights, 2) pretrained weights from ImageNet, and 3) weights pretrained on data from the Bria refugee camp in the Central African Republic). The results turn out that three-band combinations outperform RGBN bands across all types of weights and seasons. Replacing the B or G band with the N band can improve the performance in extracting dwellings during the wet season but cannot bring improvement to the dry season in general. Pretrained weights from ImageNet achieve the best performance. Weights pretrained on data from the Bria refugee camp produced the lowest IoU and Recall values.
Refugee-dwelling footprints derived from satellite imagery are beneficial for humanitarian operations. Recently, deep learning approaches have attracted much attention in this domain. However, most refugees are hosted by low- and middle-income countries where accurate label data are often unavailable. The Object-Based Image Analysis (OBIA) approach has been widely applied to this task for humanitarian operations over the last decade. However, the footprints were usually produced urgently, and thus, include delineation errors. Thus far, no research discusses whether these footprints generated by the OBIA approach (OBIA labels) can replace manually annotated labels (Manual labels) for this task. This research compares the performance of OBIA labels and Manual labels under multiple strategies by semantic segmentation. The results reveal that the OBIA labels can produce IoU values greater than 0.5, which can produce applicable results for humanitarian operations. Most falsely predicted pixels source from the boundary of the built-up structures, the occlusion of trees, and the structures with complicated ontology. In addition, we found that using a small number of Manual labels to fine-tune models initially trained with OBIA labels can outperform models trained with purely Manual labels. These findings show high values of the OBIA labels for deep-learning-based refugee-dwelling extraction tasks for future humanitarian operations.
Extracting footprints of refugee dwellings from satellite imagery supports dedicated humanitarian operations. Recently, deep-learning-based approaches have been proved to be effective for this task. However, such research is still limited due to the lack of cleaned labels for supervised-learning-based models. This research compares the performance of noisy labels from past humanitarian operations and cleaned labels by manual annotation in three classical deep learning architectures (U-Net, LinkNet and Feature Pyramid Network (FPN)) and twelve backbones (VGG16, VGG19, ResNet-18, ResNet-34, DenseNet-121, DenseNet-169, Inception-v3, InceptionResnet-v2, MobileNet-v1, MobileNet-v2, EfficientNet-B0, EfficientNet-B1). The results turn out that even though cleaned labels outperform noisy labels, the noisy labels have a high potential to replace cleaned labels because producing cleaned labels requires much more time, and predicted footprints of models trained with noisy labels are promising in humanitarian applications. Besides, the performance of the selected architectures and backbones is similar in general. Overall, FPN with VGG16, LinkNet with DenseNet-121, and U-Net with EfficientNet-B0 outperform other combinations. If considering both accuracy and training time, U-Net with VGG16 and LinkNet with ResNet-18 could be two alternatives for future research.
The Boko Haram Insurgency, a religious conflict in northeastern Nigeria, has caused around 350000 deaths and destroyed numerous populated sites in the past decade. This research used Sentinel-2 data and point data of populated sites with their status information (Functional or Destroyed) collected by Copernicus Emergency Management Service as input to train an image classification model, EfficientNet, to detect the status of the populated sites in this region. The research tested the influence of two hyperparameters (patchsize and the number of neurons in a fully connected layer) and two types of preprocessed imagery data to explore the potential of this approach for this detecting task. The results turn out the developed approach produces the highest F1-score of 0.8197 for destroyed populated sites when the model is trained with NDVI data, the patchsize is 1100m, and the number of FC neurons is 512. The results prove this approach has a high potential for long-term peace monitoring and humanitarian operations in this region.
Refugee-dwelling footprints derived from satellite imagery are beneficial for humanitarian operations. Recently, deep learning approaches have attracted much attention in this domain. However, most refugees are hosted by low- and middle-income countries where accurate label data are often unavailable. The Object-Based Image Analysis (OBIA) approach has been widely applied to this task for humanitarian operations over the last decade. However, the footprints were usually produced urgently, and thus, include delineation errors. Thus far, no research discusses whether these footprints generated by the OBIA approach (OBIA labels) can replace manually annotated labels (Manual labels) for this task. This research compares the performance of OBIA labels and Manual labels under multiple strategies by semantic segmentation. The results reveal that the OBIA labels can produce IoU values greater than 0.5, which can produce applicable results for humanitarian operations. Most falsely predicted pixels source from the boundary of the built-up structures, the occlusion of trees, and the structures with complicated ontology. In addition, we found that using a small number of Manual labels to fine-tune models initially trained with OBIA labels can outperform models trained with purely Manual labels. These findings show high values of the OBIA labels for deep-learning-based refugee-dwelling extraction tasks for future humanitarian operations.
Earth-observation-based mapping plays a critical role in humanitarian responses by providing timely and accurate information in inaccessible areas, or in situations where frequent updates and monitoring are required, such as in internally displaced population (IDP)/refugee settlements. Manual information extraction pipelines are slow and resource inefficient. Advances in deep learning, especially convolutional neural networks (CNNs), are providing state-of-the-art possibilities for automation in information extraction. This study investigates a deep convolutional neural network-based Mask R-CNN model for dwelling extractions in IDP/refugee settlements. The study uses a time series of very high-resolution satellite images from WorldView-2 and WorldView-3. The model was trained with transfer learning through domain adaptation from nonremote sensing tasks. The capability of a model trained on historical images to detect dwelling features on completely unseen newly obtained images through temporal transfer was investigated. The results show that transfer learning provides better performance than training the model from scratch, with an MIoU range of 4.5 to 15.3%, and a range of 18.6 to 25.6% for the overall quality of the extracted dwellings, which varied on the bases of the source of the pretrained weight and the input image. Once it was trained on historical images, the model achieved 62.9, 89.3, and 77% for the object-based mean intersection over union (MIoU), completeness, and quality metrics, respectively, on completely unseen images.
Shifting from effect-oriented toward cause-oriented and systemic approaches in sustainable climate change adaptation requires a solid understanding of the climate-related and societal causes behind climate risks. Thus, capturing, systemizing, and prioritizing factors contributing to climate risks are essential for developing cause-oriented climate risk and vulnerability assessments (CRVA). Impact chains (IC) are conceptual models used to capture hazard, vulnerability, and exposure factors that lead to a specific risk. IC modeling includes a participatory stakeholder phase and an operational quantification phase. Although ICs are widely implemented to systematically capture risk processes, they still show methodological gaps concerning, for example, the integration of dynamic feedback or balanced stakeholder involvement. Such gaps usually only become apparent in practical applications, and there is currently no systematic perspective on common challenges and methodological needs. Therefore, we reviewed 47 articles applying IC and similar CRVA methods that consider the cause–effect dynamics governing risk. We provide an overview of common challenges and opportunities as a roadmap for future improvements. We conclude that IC should move from a linear-like to an impact web–like representation of risk to integrate cause–effect dynamics. Qualitative approaches are based on significant stakeholder involvement to capture expert-, place-, and context-specific knowledge. The integration of IC into quantifiable, executable models is still highly underexplored because of a limited understanding of systems, data, evaluation options, and other uncertainties. Ultimately, using IC to capture the underlying complex processes behind risk supports effective, long-term, and sustainable climate change adaptation.
Current scientific discourse on the assessment of loss and damage from climate change focuses primarily on what is straightforwardly quantifiable, such as monetary value, numbers of casualties, or destroyed homes. However, the range of possible harms induced by climate change is much broader, particularly as regards residual risks that occur beyond limits to adaptation. In international climate policy, this has been institutionalized within the Loss and Damage discourse, which emphasizes the importance of non-economic loss and damage (NELD). Nevertheless, NELDs are often neglected in loss and damage assessments, being intangible and difficult to quantify. As a consequence, to date, no systematic concept or indicator framework exists that integrates market-based and non-market-based loss and damage. In this perspective, we suggest assessing risk of loss and damage using a climate change risk and vulnerability assessment (CRVA) framework: the Impact Chain method. This highly adaptable method has proven successful in unraveling complex risks in socio-ecological systems through a combination of engaging (political) stakeholders and performing quantitative data analysis. We suggest expanding the framework’s logic to include not only the sources but also the consequences of risk by conceptualizing loss and damage as harm to nine domains of human well-being. Our approach is consistent with the risk conceptualization by the Intergovernmental Panel on Climate Change (IPCC). Conceptualization and systematic assessment of the full spectrum of imminent loss and damage allows a more comprehensive anticipation of potential impacts on human well-being, identifying vulnerable groups and providing essential evidence for transformative and comprehensive climate risk management.
More than half of the world-population lives in urban areas, with more than 1 billion people lacking basic services and infrastructure. Spatially targeted, data-driven policies are crucial for sustainable urban planning to improve these situations and increase the resilience. Earth observation (EO) can support the process of achieving the SDGs, in particular SDG 11. Aiming at such high-level targets requires a multi-source data environment, defining and extracting suitable EO-based indicators and linking them with socio-economic or environmental data. When embedded in the context of humanitarian response, where physical access to regions is often limited while at the same time, insights on several scales of intervention are key to rapid decisions, the integration of (potentially) heterogeneous datasets requires adequate data assimilation strategies and a good understanding of data quality. This paper investigates the usability of datasets regarding technical and organisational aspects from an application-driven point of view. We suggest a protocol considering various quality dimensions to evaluate via scoring the fitness of multi-source geospatial datasets to integration. The aim is to provide a general orientation towards data assimilability in the context of deriving higher-level indicators, while specific constraints and the need to relativize may occur for concrete use case.
2021
While many studies exist to identify buildings from optical satellite images, radar-based approaches are still lacking in humanitarian contexts. This article outlines the main challenges related to scattering mechanisms returning from huts, tents, informal dwellings, and their natural surroundings, but also from geometric distortions caused by the sidelooking radar aperture. An outlook summarizes how these limitations can be overcome by image enhancement or multi-image composites, but also by advanced methods on building extraction, such as convolutional neural networks (CNNs). This article aims to stimulate scientific debate and to lay a foundation for the development of new methods.
For effective humanitarian response in refugee camps, reliable information concerning dwelling type, extent, surrounding infrastructure, and respective population size is essential. As refugee camps are inherently dynamic in nature, continuous updating and frequent monitoring is time and resource-demanding, so that automatic information extraction strategies are very useful. In this ongoing research, we used labelled data and highresolution Worldview imagery and first trained a Convolutional Neural Network-based U-net model architecture. We first trained and tested the model from scratch for Al Hol camp in Syria. We then tested the transferability of the model by testing its performance in an image of a refugee camp situated in Cameroon. We were using patch size 32, at the Syrian test site, a Mean Area Intersection Over Union (MIoU) of 0.78 and F-1 score of 0.96, while in the transfer site, MIoU of 0.69 and an F-1 score of 0.98 were achieved. Furthermore, the effect of patch size and the combination of samples from test and transfer sites are investigated.
For the generation of 3D city models from satellite stereo imagery beyond the generation of digital surface models (DSM) from stereo data the next crucial step is the separation of urban 3D objects from ground. To do this the most common method is the derivation of a so called digital terrain model (DTM) from the DSM. The DTM should ideally contain only the surface of the ground on which the urban objects are located. Since only the surface of the objects can be seen from space, sophisticated methods have to be developed to gain information of the bare ground. In this paper selected methods for the extraction of a DTM from a DSM are described and evaluated. The evaluation is done by applying the methods to synthetically generated DSMs. These synthetical DSMs are a combination of ground and typical urban objects put on top of it. The application of the DTM extraction methods should recover in turn the original ground model as good as possible. Also the sum of the obtained DTM and the profile of the urban objects should reconstruct the original DSM. The profile of the urban objects ist often referenced as normalized digital elevation model (nDEM). But in general the equation DSM = DTM + nDEM is not always valid – especially for buildings situated on the slope of a hill. If the nDEM would simply be the difference of DSM – DTM the slope of the hill – contained in the DTM – will be reflected on the roof of the buildings. So also an advanced method for derivation of the nDEM from DSM and DTM is presented and tested.
Amongst the many benefits of remote sensing techniques in disaster- or conflict-related applications, timeliness and objectivity may be the most critical assets. Recently, increasing sensor quality and data availability have shifted the attention more towards the information extraction process itself. With promising results obtained by deep learning (DL), the notion arises that DL is not agnostic to input errors or biases introduced, in particular in sample-scarce situations. The present work seeks to understand the influence of different sample quality aspects propagating through network layers in automated image analysis. In this paper, we broadly discuss the conceptualisation of such a sample database in an early stage of realisation: (1) inherited properties (quality parameters of the underlying image such as cloud cover, seasonality, etc.); (2) individual (i.e., per-sample) properties, including a. lineage and provenance, b. geometric properties (size, orientation, shape), c. spectral features (standardized colour code); (3) context-related properties (arrangement Several hundred samples collected from different camp settings were hand-selected and annotated with computed features in an initial stage. The supervised annotation routine is automated so that thousands of existing samples can be labelled with this extended feature set. This should better condition the subsequent DL tasks in a hybrid AI approach.
GIScience conference authors and researchers face the same computational reproducibility challenges as authors and researchers from other disciplines who use computers to analyse data. Here, to assess the reproducibility of GIScience research, we apply a rubric for assessing the reproducibility of 75 conference papers published at the GIScience conference series in the years 2012-2018. Since the rubric and process were previously applied to the publications of the AGILE conference series, this paper itself is an attempt to replicate that analysis, however going beyond the previous work by evaluating and discussing proposed measures to improve reproducibility in the specific context of the GIScience conference series. The results of the GIScience paper assessment are in line with previous findings: although descriptions of workflows and the inclusion of the data and software suffice to explain the presented work, in most published papers they do not allow a third party to reproduce the results and findings with a reasonable effort. We summarise and adapt previous recommendations for improving this situation and propose the GIScience community to start a broad discussion on the reusability, quality, and openness of its research. Further, we critically reflect on the process of assessing paper reproducibility, and provide suggestions for improving future assessments.
Sanitation refers to the provision of facilities and services for the safe disposal of human excrements and is included in SDG 6 “Ensuring the availability and sustainable management of water and sanitation for all”. With more than 2.5 billion people without appropriate facilities mainly living in cities, urban sanitation is critical to avoid harmful effects on human health and environment. Largely relying on onsite sanitation, the coordination of safe collection, treatment and disposal is essential. This requires reliable, up-to-date information that can be provided by EO data through its wide availability and objectivity. This is of particular importance in areas difficult to access and large, growing cities with urban sprawl. To facilitate the planning of faecal sludge management (FSM), priority areas where improvements are urgently needed can be identified using EO-based indicators, additionally enabling regular monitoring, evaluation of measures and targeted in-situ sampling.
Within the constraints of operational work supporting humanitarian organizations in their response to the Covid-19 pandemic, we conducted building extraction for Khartoum, Sudan. We extracted approximately 1.2 million dwellings and buildings, using a Mask R-CNN deep learning approach from a Pléiades very high-resolution satellite image with 0.5 m pixel resolution. Starting from an untrained network, we digitized a few hundred samples and iteratively increased the number of samples by validating initial classification results and adding them to the sample collection. We were able to strike a balance between the need for timely information and the accuracy of the result by combining the output from three different models, each aiming at distinctive types of buildings, in a post-processing workflow. We obtained a recall of 0.78, precision of 0.77 and F1 score of 0.78, and were able to deliver first results in only 10 days after the initial request. The procedure shows the great potential of convolutional neural network frameworks in combination with GIS routines for dwelling extraction even in an operational setting.