When it comes to Earth observations in AI for disaster risk reduction, is it feast or famine? A topical review

Earth observations (EOs) have successfully been used to train artificial intelligence (AI)-based models in the field of disaster risk reduction (DRR) contributing to tools such as disaster early warning systems. Given the number of in situ and remote (e.g. radiosonde/satellite) monitoring devices, there is a common perception that there are no limits to the availability of EO for immediate use in such AI-based models. However, a mere fraction of EO is actually being used in this way. This topical review draws on use cases, workshop presentations, literature, and consultation with experts from key institutes to explore reasons for this discrepancy. Specifically, it evaluates the types of EO needed to train AI-based models for DRR applications and identifies the main characteristics, possible challenges, and innovative solutions for EO. Finally, it suggests ways to make EO more user ready and to facilitate its uptake in AI for DRR and beyond.


Background
The new Intergovernmental Panel on Climate Change sixth assessment synthesis report serves as an important reminder that the threat of hydrometeorological and climatological extreme events grows in concert with the increase of global temperatures 10 .This situation can be attributed to the complex interactions of nonlinearities, interconnected processes, and scale effects inherent in natural phenomena.Despite a decade of advances and progresses, key open issues still remain.For instance, the Navier-Stokes equation systems in fluid dynamics remain one of the few unsolved mathematical challenges.Current modeling approaches need constant verification and 10 www.ipcc.ch.
adjustment to be deemed liable, and deterministic and stochastic models must be combined with the support of innovative approaches.After many years of exploration in fields such as chaotic systems, fuzzy logic, and fractals, it is incumbent upon the geoscience (including environmental science) and disaster risk reduction (DRR) communities to explore additional tools that can enhance our resilience to such events.One of these tools is artificial intelligence (AI).
Through their ability to efficiently digest copious amounts of data and assist with tasks such as pattern recognition and classification, AI-based models have shown promise for many research applications [e.g.health (Wiegand et al 2019), agriculture (Zhang et al 2021), or logistics (Toorajipour et al 2021)].In the fields of hydrometeorology and climatology, statistical methods for data assimilation have been adapted to integrate AI-based methods with benefits including: a more complete estimation of uncertainties (Irrgang et al 2020), a greater efficiency of established forecasting and analysis methods (see, e.g.Bauer et al 2021, Dueben et al 2022, Bi et al 2023, Ebert-Uphoff and Hilburn 2023), and an opportunity for novel combinations of data (including those derived from new observational tools such as lowcost sensors and social media, as well as Internet of Things) that can lead to tailored forecasting products (WMO 2023).In the field of DRR, AI is contributing to (near-)real-time detection and forecasting of floods (Tiggeloven et al 2021); (near-)real-time detection of wildfires (Sousa et al 2020a); (near-)realtime detection and forecasting of volcano-seismic events (Cortés et al 2021); and ex-post impact assessment (Gazzea et al 2022), among other applications.
These success stories, combined with the 2020 State of the Satellite Industry report's claim that 664 remote sensing/observation satellites transmitted Earth observations (EO) to ground stations in 2019 (Wang and Yan 2020), might give the impression that there are no limits to the availability of ready-to-use EO for AI-based models.However, the uptake of EO in traditional Earth system modelsthose models that simulate all relevant components of the Earth system, provide insight into the complex and interacting Earth processes under a changing climate, and underlie many policy-level decisionsis currently limited.For instance, McGovern et al (2020) note that less than 5% of EO are currently being used for numerical models of the Earth system.A similar phenomenon is seen in the field of AI.This begs the questions: why the disparity and, when it comes to EO in AI for DRR, is it feast or famine?
Current studies (e.g.Chi et al 2016) provide insight into the general challenges presented by EO.However, this topical review takes a novel perspective by looking at the data requirements of AI in conjunction with the unique characteristics of natural hazards.To acquire this perspective, the topical review deviates from the traditional approach used in literature-based reviews by expanding to include information derived from use cases, workshop presentations, and consultation with experts in this domain.This approach reveals underlying reasons for this disparity and innovative solutions to facilitate the uptake of EO in AI.

Data hungry models
AI-based algorithms are trained and tested using data; hence, they are often referred to as datadriven models.Rather than basing predictions on physics theories and mathematical equations (e.g.traditional flood models or numerical weather prediction models), these models learn from huge amounts of data.Therefore, it is important that the training data are representative (i.e. containing enough examples of the phenomenon to be predicted including extreme values), lacking spatial or temporal bias, reliably labeled (as needed), accompanied by metadata, and resilient to adversaries (including the failure of observational systems during an extreme event) (McGovern et al 2022).In other words, the performance of these models is predicated on having a sufficient quantity and quality of data.To ensure quality, the data need to be preprocessed and put in a format that can be digested by AI-based models.The type of data required and its availability, however, depends on the research question being addressed.Specifically, it depends on the type of natural hazard being explored and the desired output of the AI-based model.Therefore, considerations include: the geographical region(s) where a natural hazard typically occurs, logistical challenges to monitoring the region(s), the spatial scale on which an event develops and occurs (consider, e.g. the spatial granularity of data needed to capture an avalanche vs a tsunami), the time scale on which an event occurs (consider, e.g. the temporal coverage needed to record the lifecycle of an earthquake vs a flash flood), the overall aim of the AI-based model [e.g. to attribute extreme events to climate change 11,12 , or to improve volcano-seismic forecasts (Cortés et al 2021)], and the learning algorithm(s) best suited to address this research question.For instance, supervised, unsupervised, and reinforcement models each have unique data-related requirements (e.g.annotated data are required for supervised models).In addition, given the temporal dimension and spatial dependencies inherent in many variables (e.g.temperature), it is important to consider whether the model architecture can capture its related dynamics and effectively deal with possible non-or weak stationarities (Cheng et al 2015).Furthermore, imbalances in data should be considered when designing and implementing the chosen approach (Krawczyk 2016).
It is also important to note that data handling challenges, which have intensified with the emergence of new observational tools that can produce a greater heterogeneity of data types and formats (e.g.Alpert et al 2016), are not limited to the training and testing phases.They can emerge at all stages of the processing chain.Adapting to these new data handling challenges should include careful consideration of the findability, accessibility, interoperability, and reusability (FAIR) guiding principles (Fair Principles 2022).Furthermore, it is important to consider how the data produced by AI-based models can best serve downstream applications (WMO Research Board Task Team on Exascale Computing, Data Handling and AI 2023).
This topical review looks at distinguishing features of the two principle types of EO available for AI-based models in DRR applications: in-situ instrumental and remotely sensed, with a focus on satellitederived data for the latter.It also looks at an example [related to Global Navigation Satellite System (GNSS) data for tsunami detection] where both types of EO are used in conjunction.Such approaches that combine remotely sensed data with in-situ data for ground truthing can leverage the advantages of both types of EO (Albayrak et al 2013).However, they are not the focus of this paper.Rather, this topical review highlights characteristics, strengths, and selected challenges of the two types of EO, as well as creative approaches to overcoming some of these challenges.

Methods
The nexus of AI and DRR is highly dynamic.Therefore, four sources of information were used to acquire a snapshot of the current state of the art: (a) curated use cases, (b) workshop presentations, (c) scientific literature, and (d) consultation with experts from selected institutes to learn about relevant activities.Among these, (a), (b), and (d) were acquired within the framework of the United Nations (UN) Focus Group on AI for Natural Disaster Management (FG-AI4NDM) 13 , a partnership between the International Telecommunication Union (ITU), World Meteorological Organization (WMO), and UN Environment.
To curate use cases, an open call for proposals was launched on the FG-AI4NDM website (and circulated via liaison statements, the newsletters of geoscience organizations, and social media) in advance of six focus group meetings (16-17 March 2021, 24-25 June 2021, 31 August-2 September 2021, 20 October 2021, 26-28 January 2022, and 7-9 June 2022).To facilitate the systematic analysis of the proposals, proponents were provided a template that requested a project summary (including the research question and context, the method, the data, and the evaluation), a two-page project plan, a one-page outline of milestones, and a one-page description of impacts.These were presented by the proponents at the respective focus group meeting.Following a thorough discussion, the focus group decided whether to adopt the use case for inclusion in the focus group activities.This was based on criteria including: relevance (e.g.does it apply to natural-rather than man-made-hazards?), transparency (e.g. were open data used?are methods clearly described?),scientific soundness (e.g. have results been peer reviewed?does the use case build on other recognized work?), and diversity (e.g.does the use case introduce an additional natural hazard, AI-based method, or DRR product?).In total, 31 use cases were adopted14 .Many of these use cases (and related publications) are discussed in this topical review (e.g.Heck et al 2019, Sousa et al 2020a, 2020b, 2020c, Cortés et  Finally, given that many efforts related to EO in geosciences (including DRR) occur outside of academia (and, thus, might not be found in peerreviewed scientific literature), FG-AI4NDM experts from key institutes (e.g.NASA, WMO, and GEO) and the private sector (e.g.IBM, One Concern) were consulted to learn about ongoing efforts [e.g.OceanOPS, HydroHub, and the WMO Integrated Global Observing System (WIGOS)] in the domain.

Slim pickings for in-situ EO?
Long before satellites orbited and observed our planet, in-situ instrumental data (e.g.air temperature, streamflow, or seismicity) were being collected.As the instruments have evolved-both in complexity (measuring additional variables) and ease-of-operation (e.g. through automation of collection and transmission systems allowing access to real-time data; see Wood 1946)-in-situ EO have continued to be a vital source of information on our ever-changing environment.Furthermore, some key variables, such as air pressure (unless measured from a weather balloon or satellite sensor) or stream flow, are typically measured in situ.However, the spatiotemporal characteristics of in-situ EO, including continuity (or sparsity) and duration, are sometimes insufficient for AI-based methods.A recent survey 20 by the WMO reveals that two thirds of national hydrological networks are in decline and human capacity is missing in many countries.Similarly, the OceanOPS report card 21 reveals important gaps in ocean observing systems.These observational gaps need to be filled to fully benefit from AI-based methods.
In-situ measurements represent the conditions at or surrounding a collection site (e.g. at the location of a weather station or a stream gauge).Some regions subject to the same disaster risks benefit from a highdensity in-situ observation network, whereas other regions have lower spatial coverage.An example of this disparity can be seen in the distribution of avalanche observation stations in the Swiss Alps 22 (186) versus the western Himalaya (26) (Sharma and Ganju 2000).A similar imbalance can be seen in the number of seismic sensors covering Japan (200 seismographs and 600 seismic intensity meters are operated by the Japan Meteorological Agency and 3600 seismic intensity meters are managed by local governments in partnership with the National Research Institute for Earth Science and Disaster Prevention) 23 versus Haiti (ten broadband seismic stations from the Bureau of Mines and Energy; Calais et al 2019).Despite the difference in surface areas (Japan is ca.14 times the size of Haiti), the difference in sensor coverage is stark.It is noted that similar observational challenges 20 https://app.powerbi.com/view?r=eyJrIjoiMGJmMzI2MmQtZTQ2OC00NDFlLWJlNDUtZjc5NmY5OGY wNjI5IiwidCI6ImVhYTZiZTU0LTQ2ODctNDBjNC05OD I3LWMwNDRiZDhlOGQzYyIsImMiOjl9. 21www.ocean-ops.org/reportcard/. 22www.slf.ch/en/avalanche-bulletin-and-snow-situation/measured-values/description-of-automated-stations.html. 23www.jma.go.jp/jma/en/Activities/earthquake.html.1).In parallel, a company called QuakeSaver GmbH has deployed smart micro-electric-mechanical (MEM) accelerometers equipped with compute units in Japan, Germany, Switzerland, France, Montenegro, and Türkiye for monitoring earthquakes 24 .In Haiti, another approach has been used to bolster the national seismic network; personal seismometers (known as Raspberry Shakes) have been distributed among the population, providing an independent source of continuous data in case other seismic stations fall offline (Calais et al 2019).Other researchers are combining smartphone accelerometer data with artificial neural networks to distinguish earthquakes from benign vibrations (Allen et al 2020).Across our oceans-a region for which there is limited spatial coverage of in-situ EO as confirmed by the OceanOPS report card25 -submarine fiberoptical telecommunication cables are being explored by the research community for detecting earthquakes and tsunamis (Matias et al 2021).This has inspired the creation of a Joint Task Force on SMART cable systems that brings together three UN bodies: the ITU, WMO, and UN Educational, Scientific, and Cultural Organization26 .Another innovative approach to accessing oceanographic data is through bio-logging (i.e.attaching sensors to marine animals) 27 .Where these options are not feasible, some experts are using transfer learning to compensate for a sparsity of in-situ data.Here, AI-based models can be trained using data from sites with sufficient data and the model can be fine-tuned in the last layers of the deep learning model with data for another site (i.e. that would otherwise have insufficient data to train an AI-based model) (e.g.Bauer et al 2023).This method has shown effectiveness in the AI-based hazard mapping of wind-and hailstorms in Georgia (Marti 2021).
The WMO HydroHub28 encourages the uptake of such innovative approaches, including opportunistic sensors and citizen observations, to fill gaps in observational networks for hydrology.Furthermore, there is a need for innovative funding approaches to modernize and maintain measurement equipment.
Another challenge is that the temporal characteristics of in-situ EO-in terms of duration, resolution, and completeness-is highly variable (Brönnimann et al 2019).Although Nilometers already operated during Pharaonic times and the first observational network for weather (Rete Medicea) was established in 1653/4 (Camuffo 2002), national weather and hydrological services did not appear until the midnineteenth century.Furthermore, many of these hydrometeorological records are plagued by gaps and inhomogeneities (caused by, e.g.station relocation, changes in instruments, and local anthropization).
Monitoring at high-temporal resolution only became widespread or practical with the advent of automated recording in the twentieth century; although this does not include all variables.Today, WIGOS29 forms the overarching framework for the global observation of the Earth system.Counting among the most ambitious and successful instances of international collaboration of the last century, WIGOS includes surface-based observing systems with ca.11 500 land stations making at least threehourly (and often hourly) observations of meteorological parameters, as well as 1000 weather radars, 1300 upper-air stations plus about 15 ships making upperair profiles over the ocean, over 3000 automatic observing systems onboard aircrafts, 4000 routinely reporting ships, 1250 drifting buoys, more than 500 moored buoys, and many other types of observing stations (e.g.wind profilers, lightning detection systems, or tide-gauges; Barrell et al 2013).Hydrological variables are being gradually added to WIGOS as well.Moving from hydrometeorological to seismological hazards, we note that it can be exceptionally challenging to find sufficiently long in-situ records because of the timeline on which such events occur: largemagnitude earthquakes can have a recurrence period of 100-1000 yr for active plate boundaries and up to 10 000 yr for stable intra-plate regions (Kanamori and Brodsky 2004).This is considerably longer than the longest in-situ seismic records: the first seismograph appeared in 188930 , and active global positioning system monitoring of seismic zones dates only from the 1980s (Billiris et al 1991).Also in this scenario, inventive solutions can compensate to a degree.Some seismicity experts are extending the length of records through combining laboratory shear experiment ('labquake') data with numerical simulation data.By applying transfer learning, the resulting dataset can be used to predict fault-slip earthquakes (Wang et al 2021).An additional method that can be used to lengthen a dataset is producing synthetic data based on the physical properties of a natural hazard (Kuang et al 2021).Furthermore, proxy data [e.g.sedimentary seismites as indicators of past earthquakes as in Doughty et al (2014) or the use of 'trenching' through the surface expressions of faults to identify, sequence, and date past seismic events using, for example, radiocarbon dating of organic debris as in Grützner et al (2021)] can extend earthquake occurrence records by years, decades, or centuries.This approach has also been used to reconstruct other types of natural hazards [e.However, it is important that the EO capture relevant frequencies for the natural hazard of interest; therefore, attention should be paid that no critical information is lost when correcting a dataset.
Sometimes, instrumental data for the variable to be modeled are accessible, but not the metadata containing information about the collection site (e.g.topography, geology, or land use; e.g.Addor et al 2018).In other situations, instrumental data can be difficult to access due to restrictions and/or lack of user-oriented systems.For instance, researchers hoping to curate real-or near-real-time geophysical data from observatories along the ring of fire (Pallister et al 2019) note that data policies range from total sharing to total blockage of real-time data.However, many observatories also depend on external sources of data to operate.Therefore, data exchange can be mutually beneficial (Pease 2021).If, however, data cannot leave a country or observatory, distributed or federated learning are approaches that can enable models to be trained without data leaving its location (Nguyen et al 2021).
When data meet the spatiotemporal and quality requirements-and are accessible-the transmission of large data volumes can be an additional bottleneck.One bypass is computing on the cloud.According to Mei et al (2020), cloud platforms can provide better integration of geoinformation and more efficient data processing and analysis.In addition, the cloud can alleviate the hard-and software requirements on local computing devices.For example, Mendoza-Cano et al (2021) transmitted near-real-time water level, soil moisture, and meteorological data from their sensor network over a 3G network or Wi-Fi into a cloud 31 https://library.wmo.int/doc_num.php?explnum_id=10352.
environment.Other researchers favor edge computing, arguing that it can reduce data transmission, protect from network latency (or failure) issues, enable (near)-real-time data analysis, enhance data privacy, and distribute a computational workload (Tsubaki et al 2020).For example, the smart MEM accelerometers produced by QuakeSaver GmbH use a convolutional neural network to classify waveforms on the edge so that only higher-order data products are transmitted to end users32 .A similar approach is described in the crowd-sensed earthquake detection study of Bassetti et al (2022).

A bounty of remotely sensed EO?
Among the satellites collecting and transmitting remotely sensed EO, Sentinel-1, 2, and 3 produce ca.20 TB of free and open-access data daily (Esch et al 2018, Wang and Yan 2020) 33 .However, the suitability of remotely sensed EO for AI-based DRR applications depends on some of the same characteristics discussed in the previous section (figure 2).
An Earth-observing satellite can be categorized by the altitude of its orbit relative to Earth's surface as well as its inclination and eccentricity.Special orbits provide near-continuous coverage of a selected area (geostationary orbit) or global coverage, passing a point on Earth at the same local time (sun-synchronous, near-polar orbit) 34 .These orbital characteristics, in conjunction with the acquisition mode or spectral band, strongly impact the spatiotemporal coverage and resolution (i.e.spatial, temporal, spectral, or radiometric) of the resulting EO.On board the satellites, passive (e.g.radiometer) and active (e.g.scatterometer) sensors record the Earth's surface and atmosphere as optical or synthetic aperture radar (SAR) imagery, respectively.Although this section focuses on imagery, satellites also produce data in other modalities with relevance for DRR.For instance, GNSS satellites can provide data on the rotation of the Earth, variations in the hydrosphere, and characteristics of the Earth's ionosphere and troposphere 35 .GNSS satellites have also been used to estimate near-surface soil moisture (e.g.Kim and Lakshmi 2018).
For some purposes, the spatial resolution of satellite-derived EO can pose a challenge.For instance, Hafner et al (2021) used Sentinel-1 (10 m × 10 m) SAR imagery, Sentinel-2 (10 m × 10 m) optical imagery, and higher-resolution optical imagery (1.5 m × 1.5 m) from the commercial low-Earth-orbiting and sun-synchronous SPOT6 satellite to develop snow avalanche maps.Although SPOT6 outperformed Sentinel-1 and 2 at detecting the largest avalanches, the authors found that all three posed issues for mapping avalanches of smaller sizes.Kovács et al (2022) had a similar challenge when attempting to use Sentinel-1 SAR imagery to detect landslides.Lato et al (2012), meanwhile, eschewed free Sentinel data and opted instead for commercial optical imagery from the QuickBird-2 satellite.This low-Earth-orbiting and sun-synchronous satellite boasts a spatial resolution from 0.61 m × 0.61 m to 0.72 m × 0.72 m (for panchromatic imagery) and from 2.4 m × 2.4 m to 2.6 m × 2.6 m (for multispectral imagery) over a 16.5 km wide swath 36 .Through segmentation and classification, the authors could automate the detection and mapping of snow avalanche deposits.Unlike spatial resolution, coverage has rarely been identified as a limitation when using satellite-derived EO (e.g.passive microwave data) to monitor snow avalanches.This is because 36 https://earth.esa.int/eogateway/missions/quickbird-2.
global coverage is provided every several days.However, these EO have other challenges as described below.
When monitoring and detecting natural hazards on relatively short time scales (e.g.minutes vs days), the temporal resolution can be the bottleneck.For example, Sousa et al (2020b) found that the wait interval for overpasses made most satellite data insufficient for fire detection.Therefore, they used thermal images captured by mobile robots-an alternative source of remotely sensed data.Sometimes, the record duration is the dominant issue.For instance, Wimmers et al (2019) used satellite microwave imagery from several low-Earth-orbiting satellites (including Tropical Rainfall Measuring Mission and Aqua) to estimate tropical cyclone intensity using a convolutional neural network.They found that the model performed poorly when estimating the intensity of category five tropical cyclones, which they ascribed to a shortage of data for this category.One option to augment such a dataset is through generating synthetic tropical cyclone data as described in Nederhoff et al (2021).
When it comes to the quality of remotely sensed EO imagery, some persistent issues include distortion, noise, striping, illumination, and site obscuration.Distortion can be ascribed to topographical features and/or the satellite viewing angle.It can usually be remedied via orthorectification.Here, surveyed control points are used to calibrate an image 37 .Speckle noise, which is common on SAR imagery, results when environmental conditions influence the sensor during image acquisition.These blemishes can interfere with image interpretation, classification, and segmentation to varying extents.For instance, Boonprong et al (2018) noted that some supervised machine learning (ML) methods (whereby a model is trained to learn the relationship between data features and labels) are more robust to speckle than others.Fortunately for the lessrobust ML methods, many studies have evaluated speckle reduction methods (e.g.genetic algorithms) and image restoration (e.g.Maity et al 2015).Another important consideration when using SAR is the impact of radio frequency interference on data quality.Such interference-which is exacerbated in certain geographical regions (e.g.East Asia, Europe) and for low frequency bands-occurs when signals emitted from other radiation sources (e.g.mobile wireless technologies) occur in the same radio frequency band as the SAR.However, advanced signal processing techniques can provide some reprieve (see, e.g.Tao et al 2019).In optical imagery, another artifact is characterized by columns or rows appearing brighter or darker than those adjacent.These stripes, which can blur features in an image (Li et al 2022), can be caused by the calibration and performance of different detectors within a sensor, the reflectance on opposite sides of the rotating mirror that redirects incoming light into the sensor's optical path, and distinctions in the solar-sensor geometry in the detectors that scan a target object in a slightly offset way (Mikelsons et al 2014).To detect and remove stripes, methods can leverage the periodic repetition of stripes and changes in brightness relative to nonstriped portions of an image (Li et al 2022).These include low-pass filtering in the frequency domain through the discrete Fourier transform (Shen and Zhang 2009), splitting an image into striped and unstriped components and then removing stripes from the former via a nonlinear filter (Mikelsons et al 2014), and applying the maximum a posteriori algorithm (Shen and Zhang 2009).Another challenge in optical imagery stems from differences in illumination, shading, and reflectance.These can be influenced by topography and/or land cover as 37 https://earthobservatory.nasa.gov/features/GlobalLandSurvey/page3.php.
well as the sun's altitude and azimuth.Shadows, for instance, have been blamed for a neural network's misclassification of snow-covered areas in Cannistra et al (2021).In the Zhao et al (2018) study of landslides in EO from bands 1-7 in the Landsat-8 Operational Land Imager, specimens in orographic shadows were removed from the dataset because of differences in the spectral reflectance.Illumination can also pose a challenge when using satellite-derived EO for avalanche mapping.For instance, Hafner et al (2021) found that considerably fewer avalanches could be detected in optical data from SPOT6 under shady versus fully illuminated conditions.Fortunately, this is an active area of research with methods being developed to detect and correct such flaws through, for instance, using stereo-mapping techniques to better understand the topography of the study region (e.g.Shedlovska and Hnatushenko 2016, Barbarella et al 2017, Zhou et al 2021).Another important issue is the obscuration of features in optical imagery.For instance, some researchers elect not to use optical imagery due to the risk of obscuration by cloud cover.Instead, they turn to SAR, which operates at wavelengths that are resilient to cloud cover (Deijns et al 2022) 38 .However, researchers might prefer optical imagery to SAR if, for instance, a longer historical record is needed.In Meraner et al (2020), a deep residual neural network and SAR-optical fusion were used to leverage the benefits of both sensors and to remove clouds from Sentinel-2 data.In the FloodSENS project, which is led by RSS-Hydro and the European Space Agency's InCubed program, an automated supervised ML-based flood mapping algorithm is being used to reconstruct flood extents in cloud-obscured optical imagery.Thus far, the algorithm has been tested for Mozambique and Hungary with promising results39 .It is worth considering, however, that while satellite-derived images might need a certain level of image calibration or correction to improve quality, these measures should not compromise the robustness or generalizability of the resulting model.
Data annotation is a prerequisite when applying supervised ML methods for in-situ instrumental and remotely sensed data (e.g.Boonprong et al 2018).It is addressed in this section because it can be especially challenging for remotely sensed EO due to the high volumes of data.In a hurricane study by Kuzin et al (2021), crowd sourcing was used to generate pointbased annotations of damage for a set of satellite optical images.After quality controlling these annotations, the authors trained a faster region-based convolutional neural network (R-CNN) to automate the annotation process for additional images (figure 3).In Sousa et al (2020c), a lack of large-scale wildfire databases motivated a different human-in-theloop system, which integrates domain expertise in a semi-automated ML-based data annotation pipeline.Another human-in-the-loop system appears in the natural hazard-versatile, no-code, and open-source Curator tool from Venguswamy et al (2021).First, the tool trains a self-supervised algorithm on a random subset of unlabeled satellite optical images.Then, researchers supply a reference image of interest and similar images are extracted from the dataset.The resulting images are manually annotated and used to train a classifier, which is able to find other similar images for human inspection via active learning.Meanwhile, Gazzea et al (2022) circumvented the data annotation issue by using a segmentation model, spectral vegetation index, and variational autoencoders to automatically detect fallen trees in preand post-event satellite optical images with a limited training set.
Also in the case of remotely sensed data, access can be a hurdle.Sometimes, it is restricted to paying customers because the data are privately owned (e.g.those from QuickBird-2; Lato et al 2012).During a major disaster, however, these costs can be waived if an authorized user activates the International Charter.For instance, following a landslide in Slovenia during November 2000, before and after imagery from Airbus' SPOT4 satellite were made available to civil protection 40 .Other times, data export limitations or other protections can prevent the transfer of data.For instance, natural hazards such as tsunamis, earthquakes, and volcanic 40 www.airbus.com/en/newsroom/stories/2020-11-getting-lifesaving-satellite-imagery-to-where-its-needed-most.
eruptions (Matoza et al 2022) generate atmospheric waves that can be detected using GNSS observations (Komjathy et al 2016), which require satellites and continuously operating ground stations around the world (Vergados et al 2020b).The application of advanced GNSS real-time processing for positioning and ionospheric imaging (Meng et al 2019) provides significant improvements to tsunami early warning.However, some countries and territories do not readily share real-time GNSS observations (Vergados et al 2020a, Martire et al 2021), either due to political prohibitions or internet bandwidth and connectivity restrictions.Research exploring the feasibility of using AI for decentralized domestic processing of geodetic data is currently underway within the International Association of Geodesy, with the goal of enabling sharing of life-saving geodetic realtime disaster information within the parameters of data export restrictions.
As shown in the tsunami example, some satellite EO are electronically transmitted directly to a ground station.Other times, EO can be temporarily stored onboard a satellite or pass through the Tracking and Data Relay Satellite System until reaching the line-ofsight of another ground station 41 .However, petabytescale volumes of remotely sensed data can pose a problem for many data centers.Thus, some experts have turned to cloud platforms.For instance, Ilmy et al (2020) used the Google Earth Engine to derive an ML-based landslide susceptibility map for East Java, Indonesia.Mayer et al (2021) leveraged the cloudbased Google AI Platform and Google Earth Engine, which were recently integrated, to automate deep learning-based surface water mapping for Cambodia using Sentinel-1 SAR data.Other experts prefer edge computing-whereby AI algorithms are trained on board a satellite and only second-order products are transmitted (thus reducing bandwidth requirements and having lower latency than cloud computing)-as an attractive alternative to centralized or cloud computing (e.g.Giuffrida et al 2022).

Discussion and next steps
At COP27 in November 2022, the WMO launched an Executive Action Plan-Early Warnings for All 42 .In the context of DRR (and with applicability for other fields that leverage EO; e.g.geosciences, agriculture, energy, logistics, and insurance/reinsurance), key barriers to the uptake of EO in AI-based models-including those contributing to early warning systems-are related to data properties.The newly approved WMO Unified Data Policy for the whole 41 www.nrcan.gc.ca/maps-tools-and-publications/satelliteimagery-and-air-photos/tutorial-fundamentals-remotesensing/satellites-and-sensors/data-reception-transmissionand-processing/9327. 42https://library.wmo.int/doc_num.php?explnum_id=11426.
Earth system will substantially contribute to solving some of those issues 43 .For some projects, innovative solutions have proven successful at overcoming these barriers.However, to upscale the use of EO in AIbased models for DRR, deliberate steps can be taken by our community.
First, we can systematically identify the properties that are required from EO.Here, the FG-AI4NDM is making strides by producing a detailed technical report that deconstructs EO properties and data handling methods for the 31 use cases and proposing standards.This report considers major UN initiatives, including the Sustainable Development Goals 44 , Sendai Framework for Disaster Risk Reduction 2015-2030 45 , and the UN International Geospatial Information Framework 46 to harmonize terminology, strategic objectives, and operational planning with the UN Committee of Experts on Global Geospatial Information Management (UN GGIM; Craddock et al 2021).This will enable integration, visibility, and uptake of such proposed standards in strategic regional, national, and global geospatial information management arrangements 47 .Recently, the Group on Earth Observations (GEOs) has also begun to liaise with the standardization initiative ISO/TC211 48 .
To drive uptake, we need to facilitate and encourage data sharing through targeted policy (e.g. the aforementioned WMO Unified Data Policy) and a demonstration of benefits.GEO has been a strong advocate for broad and open data-sharing policies and practices through its Data Sharing and Management Principles 49 , which have helped member countries and organizations to evolve from restricted data policies to open data approaches.Although GEO has not yet focused on open data sharing and management specifically for AI, this could be an area to further build on its work in collaboration with the aforementioned focus group.Data sharing can also be cultivated through platforms such as the GEO Knowledge Hub 50 , Copernicus Open Access Hub, the NASA Open Data Portal 51 , Capella Space Open Data Gallery 52 , SpaceNet 53 , WorldStrat Database 54 , and BigEarthNet 55 .The Land, Atmosphere Near Real-Time Capability for EOS56 , which is hosted by NASA, has already made headway in curating and sharing EO data from ten instruments on board Terra, Aqua, Aura, Suomi NPP, GCOM-W1, and the ISS.These data are provided within 3 h of the satellite overpass, making them particularly useful for monitoring natural hazards in near-real time.In addition to sharing data, downloadable open access software and online platforms can support data processing and augmentation, algorithm training, and benchmarking.Some examples include: AGORA-EO57 , OpenEO58 , SeisBench59 , and AiTLAS60 .Furthermore, as the body of data continues to grow, we need to adapt by advancing edge-and cloud computing systems.
We believe that further advances can be substantially accelerated by bringing together open EOincluding benchmarking datasets-and tools in a centralized platform so that they can be easily accessed by users for training and testing AI-based models.This will also make it possible to take an inventory of what natural hazards are being captured in EO and in what way (e.g. at what resolutions, what quality), so that it is known where additional monitoring is required.Some attempts are already being made to call for or create so-called 'grand challenge' infrastructures that curate observations, algorithms, or other tools for a specific natural hazard type [e.g.earthquakes as in Ben-Zion et al (2022) or k-scale climate modeling as in Slingo et al (2022)] and enable testing of models.However, we urge a more holistic and inclusive approach, in which EO data, algorithms, and tools for the spectrum of natural hazard types are curated to enable exploration of interactions such as compound and cascading events, tipping points, and threshold effects.This will also enable us to test the limits of AI for the most elusive types of extreme events such as black swan phenomenon that are not represented within our body of data.To define, produce, and maintain such a global and multi-hazard platform inevitably requires substantial international and interdisciplinary collaboration-bringing together experts in natural hazards and the broader geosciences, in EO, in DRR, and in ML/AI [as described in Kuglitsch et al (2022b)]-along with private sector expertise.
Finally, we note that-in addition to those related to data-other decisions can impact the effectiveness of AI technology in DRR.These decisions can occur at the problem definition stage, model training and testing stage, or (if relevant) operational implementation stage (Kuglitsch et al 2022a).

Figure 1 .
Figure 1.A seismic and infrasound avalanche detection system is powered by solar panels in the Dischma Valley, Switzerland.In the background are scars from an avalanche that released several days prior.Photo credit: Alec van Herwijnen, SLF.
g. sedimentary turbidites as indicators of past floods or landslides (Stewart et al 2011), tree rings as records of past droughts or debris flows (Toreti et al 2013, Franco-Ramos et al 2019), and speleothems as archives of past wildfires (McDonough et al 2022)].However, the temporal resolution of proxy records depends on the nature of the archive with considerably lower certainty than direct observations.The quality of in-situ data is another important characteristic.It can be impacted by many variables including the location of the collection sites (and nearby land use and land use change) and problems related to instrumentation (Kuglitsch et al 2012).When detecting extremes or atypical behaviors, quality-related issues can mean the difference between achieving a world record maximum temperature (or not; El Fadli et al 2013).Fortunately, several methods can provide insight into the quality of in-situ EO (World Meteorological Organization 2020).These include: collecting, standardizing, and managing metadata, which provide critical information about the measurement methods, conditions, and history; running standard quality control tests (e.g.Durre et al 2010, Toreti et al 2022); and applying statistical break detection tests(Caussinus and Mestre 2004, Kuglitsch et al 2012, Lindau and Venema  2018).If needed, and when it is feasible, inhomogeneities can be corrected via data homogenization methods(Kuglitsch et al 2009, Toreti et al 2011, 2012)  31 .

Figure 2 .
Figure 2. (left) Traditional methods and (right) challenges to gathering EO from (top) remote sensing and (bottom) in-situ devices (credit: Ashley Nilo and Allison Craddock, NASA Jet Propulsion Laboratory / California Institute of Technology).

Figure 3 .
Figure 3. High-resolution satellite-derived EO from Maxar are annotated using a combination of crowdsourcing (over the Zooniverse platform) and machine learning (Kuzin et al 2021).Reproduced from Kuzin et al (2021).CC BY 4.0.
Grützner et al 2021, Kovács et al 2022); general challenges and opportunities presented by EO in geosciences (including DRR; e.g.Shen and Zhang 2009, Maity et al 2015, Esch et al 2018, Wang and Yan 2020); and general challenges and opportunities presented by AI in DRR (e.g.Krawczyk 2016, McGovern et al 2020, Kuglitsch et al 2022b, WMO 2023).The last search was made in September 2022, from Europe.