Monitoring the Recovery Process After Major Hydrological Disasters with GIS, Change Detection and Open and Free Multi-Sensor Satellite Imagery: Demonstration in Haiti After Hurricane Matthew

Abstract

Recovery from disasters is the complex process requiring coordinated measures to restore infrastructure, services and quality of life. While remote sensing is a well-established means for damage assessment, so far very few studies have shown how satellite imagery can be used by technical officers of affected countries to provide crucial, up-to-date information to monitor the reconstruction progress and natural restoration. To address this gap, the present study proposes a multi-temporal observatory method relying on GIS, change detection techniques and open and free multi-sensor satellite imagery to generate thematic maps documenting, over time, the impact and recovery from hydrological disasters such as hurricanes, tropical storms and induced flooding. The demonstration is carried out with regard to Hurricane Matthew, which struck Haiti in October 2016 and triggered a humanitarian crisis in the Sud and Grand’Anse regions. Synthetic Aperture Radar (SAR) amplitude change detection techniques were applied to pre-, cross- and post-disaster Sentinel-1 image pairs from August 2016 to September 2020, while optical Sentinel-2 images were used for verification and land cover classification. With regard to inundated areas, the analysis allowed us to determine the needed time for water recession and rural plain areas to be reclaimed for agricultural exploitation. With regard to buildings, the cities of Jérémie and Les Cayes were not only the most impacted areas, but also were those where most reconstruction efforts were made. However, some instances of new settlements located in at-risk zones, and thus being susceptible to future hurricanes, were found. This result suggests that the thematic maps can support policy-makers and regulators in reducing risk and making the reconstruction more resilient. Finally, to evaluate the replicability of the proposed method, an example at a country-scale is discussed with regard to the June 2023 flooding event.

1. Introduction

According to the United Nations Office for Disaster Risk Reduction (UNDRR) [1], the Disaster Risk Management Cycle encompasses different phases, i.e., Response, Rehabilitation and Recovery, Prevention and Mitigation and Preparedness. This definition is very consolidated across the international community. However, not all the above phases are given the same level of attention, efforts and investments. This appears to be the case for Rehabilitation and Recovery, and, in particular, there are fewer initiatives aiming to monitor if a territory hit by a disaster effectively recovers, how long this process takes and if it happens in a way that improves resilience to future similar events.

In the literature, the concept of recovery from disasters and natural hazards is defined as the complex process that requires careful planning and implementation of a range of measures to restore infrastructure, basic services and the quality of life of affected residents [2]. Remote sensing data and geospatial analyses are established means for impact and damage assessment in the aftermath of a disaster, though they are not yet part of a common practice for recovery monitoring. Satellite imagery can indeed play a crucial role in this process, providing important and up-to-date information on the progress of reconstruction and changes in the affected environment. During the recovery process, satellite imagery can be used to provide regular updates on reconstruction progress and highlight changes that are related to the natural and human-induced recovery of the territory. This includes identifying newly built structures, assessing progress in the rehabilitation of critical infrastructure, such as roads, bridges and utilities, and monitoring the restoration of the natural environment, such as reforestation, rehabilitation of coastal areas and the appropriate use of land affected by the natural disaster.

The recovery period is typically divided into two distinct phases in the temporal development of a disaster, with satellite data playing different roles with respect to them. Figure 1 shows the conceptual sketch developed by the Committee of Earth Observation Satellites (CEOS) Working Group on Disasters (WGD). According to this sketch, the demand for geoinformation useful for addressing both the emergency and post-emergency phases varies depending on the stage we are in [2]. During a crisis, there is the highest current use of geoinformation, given that one of the key needs is to access a basemap showing the impact of the event in order to assess damage and arrange for civil protection measures. As we move into the post-crisis recovery phase, the demand for geoinformation changes from basemaps to more complex information, enabling assessment and long-term planning. Currently, satellite data are predominantly used during the emergency phase and less so to monitor the whole recovery process. A further challenge is represented by the fact that, while during emergencies humanitarian aid typically involves geospatial organisations providing the needed basemaps in a very timely framework, soon after the emergency phase is over, such costless support becomes less frequent, and thus the affected countries are unable to access the needed geoinformation to support recovery monitoring. The most common causes include (but are not limited to) a lack of trained technical operators and data analysists, a lack of funding resources to access data (either satellite or more generally geospatial data to perform surveys) and a lack of computing facilities.

This is a gap that CEOS WGD has recently recognised (see the concept of Generic Recovery Observatory; [3]), and growing efforts are made to promote the utilisation of Earth observation data for this task. In this context, this study aims to propose a methodology based on GIS and open and free multi-sensor satellite imagery which is both feasible and manageable for end-users, with a particular care for those who are non-specialists, in order to ensure a long-term sustainability in the use of satellite technologies.

The disaster event that was selected to demonstrate the proposed methodology is Hurricane Matthew, which happened in September–October 2016. The 2016 hurricane season was one of the most intense recorded in the Atlantic Ocean, and its most powerful category 5 hurricane, Hurricane Matthew, triggered a major catastrophe across the island of Haiti. The hurricane caused a significant amount of damage to infrastructure, homes and crops, leading to a humanitarian emergency in the region. In addition, the hurricane induced flooding and some landslides, resulting in a large number of deaths and displaced persons. The impacts of Hurricane Matthew on Haiti were particularly devastating due to the country’s vulnerability to extreme weather events and the lack of basic infrastructure and services to deal with the emergency. Therefore, while in the immediate aftermath priority was given to early response and humanitarian emergency, in the following months and years the main concern was to ensure a recovery process consisting not only of the reconstruction of damaged buildings and infrastructure, but also a reorganisation of urban settlements and the environment that would ideally decrease exposure to future events [4].

Specifically, this study examines the capability of Synthetic Aperture Radar (SAR) Sentinel-1 and optical satellite images from Sentinel-2 to detect changes caused by severe hydrometeorological events following a natural disaster, such as the one that occurred in Haiti, and subsequently to conduct GIS analysis using data from these satellites for monitoring the recovery process on the island. By utilising SAR amplitude and multispectral change detection methods, the study aims to comprehend the extent of damage caused by the hurricane, including the identification of the most affected areas, namely, the Grand’Anse region and the southern region, where the two major urban settlements of Jérémie and Les Cayes are located, respectively. The study analyses the spatial patterns and temporal evolution of changes in infrastructure, vegetation and flood-affected agricultural areas to gain insights into the dynamics of construction, reconstruction and the recovery process of the island. A time-series analysis approach was employed to assess differences between satellite images before and after the hurricane and in the years after, facilitating the identification of areas impacted by the natural disaster and those that exhibited reconstruction efforts and natural and/or human-induced recovery. The thematic maps presented in the paper serve as proof of concepts of the value-added geoinformation that local stakeholders and decision-makers may receive should this methodology be implemented operationally and run by local technical officers and data analysts. In this respect, the methodology is intentionally designed to rely on simple and consolidated processing techniques, as well as free access processing tools, so as to test a technical barrier-free workflow that can be effectively exploited in low-income countries and real-world scenarios where resources are scarce.

To demonstrate the replicability of the methodology and prove that it is independent of the specific characteristics of the hydraulic/weather-related event, a test is presented on a series of floodings that happened across Haiti on 2 and 3 June 2023. Although it received less attention from the media, this was an extreme precipitation event caused by a low-pressure system developed in the Caribbean, resulting in the loss of at least 42 lives and damage to over 13,600 houses, necessitating the evacuation of residents. Therefore, it acted as a suitable case for testing replicability. The paper outlines the practical advantages of the proposed methodology, as well as the limitations intrinsic to the coarse resolution and typology of input data. Finally, ways to further improve the methodology towards automation and more independence from the operator’s expertise and prior knowledge of the context are outlined.

2. Case Study

2.1. Hurricane Matthew

This tropical cyclone formed in the Atlantic Ocean on 28 September 2016 near the Leeward Islands in the Caribbean Sea. The storm first started as a tropical depression, but rapidly intensified as it advanced westwards. On 30 September, Matthew became a tropical storm and continued its west–north-westward path. On 1 October, Matthew intensified further and became a Category 5 hurricane, the highest according to the Saffir–Simpson hurricane intensity scale. The storm remained a Category 4 hurricane for several days before weakening to Category 3 and then to Category 2. However, on its way to Haiti, Hurricane Matthew regained strength and intensified to Category 4 before making landfall. Hurricane Matthew made landfall in Haiti on 4 October 2016, bringing with it strong winds of 230 km/h, torrential rains and storm waves [5]. In addition, Hurricane Matthew also affected other countries in the region, such as the Dominican Republic, Cuba and the Bahamas. The hurricane continued its trajectory northwards, hitting the east coast of the United States. The storm made landfall in North Carolina on 8 October as a Category 1 hurricane, before continuing north-eastwards and dissipating over the Atlantic on 10 October 2016.

Figure 2 shows the trajectory of Hurricane Matthew over Haiti, where it caused catastrophic damage in the study area. It is estimated that more than 2 million people were affected by the hurricane, with hundreds of thousands of homes destroyed or damaged and thousands of people displaced. According to reports, the hurricane caused more than 500 deaths and thousands of injuries [6]. The strongest winds and heaviest rainfall were recorded in this study area; the hurricane caused flash floods and landslides that destroyed houses, roads and bridges. In addition to material damage, Hurricane Matthew also had a significant impact on the country’s infrastructure, with the destruction of water and sanitation systems and the disruption of electricity and communications services. Agriculture, which is an important source of income for many people in Haiti, also suffered great losses due to the destruction of crops and the death of animals due to flooding. Hurricane Matthew was considered as one of the worst natural disasters hitting Haiti in recent decades, and its impact on the country was devastating, further exacerbating the poverty and vulnerability of this country that, from an economic point of view, is classified as the least developed in the northern hemisphere.

With these characteristics, Hurricane Matthew was a disastrous event that, unfortunately, effectively represents the category of severe hydrological events that could strike Haiti and/or other Central American countries again in the future. Therefore, the methodology developed and tested in this study has exportability potential.

2.2. Study Area

Figure 3 shows in red the whole area that was most affected by Hurricane Matthew, i.e., the departments of Grand’Anse, Nippes and Sud in south-western Haiti. This is also the area that, in the present study, was investigated to assess the impact during the emergency, i.e., by comparing pre-, cross- and immediately post-event situations. The investigation of the recovery and reconstruction period from 2017 to 2020 was undertaken over eight sectors (see black boundaries in Figure 3). These sectors were intentionally selected based on the degree of the hurricane’s impacts. Not surprisingly, these sectors are those that include the main urban centres, infrastructural assets, higher population density and tree stock, i.e., the main elements that would be at risk in the event of a hurricane and that, in the case of the 2016 event, constituted the main human, economic, social and environmental damage.

The selected sectors provide a wide range of the abovementioned impact typologies. While the cities of Jérémie to the north and Les Cayes to the south represent the most populated and built-up areas, the Troupeau sector was selected for its specific dynamics due to the temporary floods caused by the hurricane, and Macaya Park was selected for the significant damage to the local biosphere. The Miragoane and Ile-à-Vache areas were chosen for their natural mangrove land cover, which acts as an important barrier to withstand waves due to a hurricane, and Hurricane Matthew strongly affected these areas during its passage.

It is important to highlight that, although the selection of these sectors clearly stems from experimental technical–scientific choices made for conducting this research, this choice also aligns perfectly with the areas and themes that were selected, investigated and addressed by the CEOS Recovery Observatory project (https://www.recovery-observatory.org/drupal/en accessed on 8 September 2024) [7]. Therefore, this research also simulates what would be the set of scenarios that local end-users would investigate and monitor using the proposed GIS and remote sensing-based recovery assessment methodology.

3. Methodology

The methodology employed in this study is based on the analysis of free and open satellite image data collected before and soon after the hurricane, as well as every six months for subsequent years when the recovery period is supposed to happen. Furthermore, the methodology relies on the combination of optical and SAR satellite imagery. The procedure has been divided into five main phases: data collection, generation of SAR data intensity change, selection and extraction of relevant changes, generation and analysis of intensity change and the production of maps depicting the hurricane impacts and proxies suggesting anthropogenic activities related to reconstruction and recovery.

3.1. Satellite Data

The methodology was developed for use in assessing case studies related to hydrological and weather-related events such as hurricanes, storms and floods. Therefore, to account for the typical cloudiness that characterises the weather during such hazard events, the best satellite datasets are those collected by SAR sensors. Furthermore, given that the demonstration is carried out over a Caribbean country, thus in a tropical climate zone affected by high likelihood of cloudiness, the choice of SAR data is even more recommended in order to achieve a timely observation of impacts and changes at the desired epochs.

In the first phase, open and free SAR Sentinel-1 satellite images were sourced from the Copernicus programme. Of the two polarisations available, the vertical–vertical (VV) polarisation was selected given the known stronger signal compared to the other vertical–horizontal (VH) polarisation. Furthermore, the methodology exploits VV imagery collected along both ascending and descending geometries. This choice allows the analyst to benefit from the best terrain visibility conditions, which is particularly crucial in study areas that are not plain and are hilly to mountainous, such as the south-western departments of Haiti.

The optical imagery included both Copernicus Sentinel-2 scenes and higher spatial resolution images downloaded from the free-access programme from Norway’s International Climate and Forest Initiative (NICFI). Optical imagery was used not only to capture surface changes to complement observations made with SAR data, but also as a means of visual verification to better interpret the patterns and ground objects that were not clearly detectable in SAR images.

It is important to highlight that, although it is established that emergency mapping mainly uses high-resolution optical data, often provided commercially, one of the main objectives of this research is to demonstrate the benefit of free and open data, like, for example, Sentinel-1 and 2. Despite their lower spatial resolution, these data can represent an important data source for the initial emergency phase soon after a disaster event and can support cost-effective monitoring during the recovery period, especially when funding resources are limited. Indeed, this selection of data presents some limitations, which we discuss later, but it has an advantage in terms of sustainability, especially for end-users in developing countries like Haiti, who have limited access to high-resolution satellite information in a timely manner during natural emergencies.

3.1.1. 2016 Hurricane Matthew

Figure 4 shows the temporal distribution of the satellite data used for the different phases. Sentinel-1 data was selected from August to December 2016 to cover the period before, during and after the hurricane, to have a comparison of the changes produced immediately after the hurricane and to minimise the likelihood of capturing unrelated changes. Image pairs with time intervals of approximately 60, 30 and 15 days were compared to detect changes in the radar backscatter signal (amplitude), as highlighted by grey bars in Figure 4a. Starting from January 2017, image pairs were selected at 6-month intervals, thus achieving change maps between approximately the beginning and middle of each year (Figure 4b).

This methodological choice follows the rationale by which the months soon after are typically characterised by sudden and frequent changes related to emergency operations, population sheltering and initial recovery activities. Therefore, shorter time intervals are more likely to capture these dynamic changes. Reconstruction and mid- to long-term recovery are more gradual processes, and significant changes are produced over a longer time interval. In addition, the true reconstruction and recovery phase in the following months to years may occur on an urban or regional scale with a longer time span. At the resolution level of the Sentinel data and the mapping scale allowed by these data, progress, e.g., in the reconstruction of dwellings, may not be evident in change detection pairs at 6-day (thus weekly) or monthly intervals, but rather over periods of several months at a time.

The dates of the optical data (Sentinel-2 and NICFI) were selected to coincide as closely as possible with the dates of the SAR data (

Table 1. Dates of the SAR data used for the analysis.

SAR Images	Geometry	Date
Sentinel-1	Ascending	07 August 2016
		31 August 2016
		24 September 2016
		18 October 2016
		11 November 2016
		05 December 2016
		29 December 2016
		22 January 2017
		09 July 2017
		05 January 2018
		21 August 2018
		24 January 2019
		16 August 2019
		07 January 2020
		03 September 2020
	Descending	05 August 2016
		29 August 2016
		22 September 2016
		10 October 2016
		03 December 2016
		27 December 2016
		20 January 2017
		07 July 2017
		03 January 2018
		19 August 2018
		10 January 2019
		14 August 2019
		05 January 2020
		01 September 2020

and

Table 2. Dates of the optical data used for the analysis.

Optical Images	Cloud COVERAGE	Date
	Low Clouds	19 September 2016
		09 October 2016
	Mid Clouds	28 November 2016
	Low Clouds	07 January 2017
		11 July 2017
Sentinel-2		28 November 2017
	Mid Clouds	03 December 2017
		27 January 2018
		30 August 2018
	Low Clouds	17 January 2019
		21 June 2019
		12 January 2020
	Free Clouds	03 September 2020
		August–December 2016
Planet (NICFI)	Free Clouds	August–December 2018
		August–December 2019
		01–30 September 2020

); however, the cloudiness in the study area required a check against the SAR results. NICFI data with a high spatial resolution of 5 metres were acquired from the information available for free access semi-annually. These images contain a cloud-cleaned composition or mosaic of several 6-month images from the middle to the end of the year and from the beginning to the middle of the year, i.e., two for each year. The only exception is the image of September 2020, which is a composition of only one month. The creation of this clean composition from the clouds is important for the analysis and review of the data, as well as for the control of outliers in the SAR data.

3.1.2. June 2023 Event

To demonstrate the replicability of the methodology, the flooding event that occurred north and south of Haiti’s capital city in June 2023 was selected.

To observe the changes that occurred after the extreme precipitation, Sentinel-1 images collected on 27 May (pre-event) and 8 June (post-event) 2023, both in ascending orbit and with VV polarisation, were chosen. Figure 5 shows the area affected by the event and the satellite data coverage. As can be seen in the terrain elevation model, the detailed analysis was carried out predominantly on flat land areas that are prone to flooding after hydraulic and weather events.

3.2. Ancillary Geospatial Data

In addition to satellite data, ancillary vector data were accessed to help in the interpretation of SAR-based change patterns and their conversion into geospatial units pertaining to elements at risk.

Building data in Haiti from 2020 were accessed from Open Street Map (OSM). This dataset shows objects that represent completed constructions, like houses, offices, schools, hospitals and others. These are typically tagged as “building”, or according to more specific categories like “house” or “school” [8].

Furthermore, geospatial information on the damage caused by the hurricane and the subsequent recovery was accessed through the activation EMSN050 managed by the Copernicus Emergency Management Service (EMSN). This dataset shows the buildings and infrastructure in specific areas of Les Cayes and Jérémie, and also of temporary camps after the disaster [9].

The geographic database of Haiti’s land cover for the year 2017 (OCS_Haiti_2017) was made available to this research via the CEOS WGD, courtesy of the National Centre for Space Studies (CNES) that developed it based on Sentinel-2 data and shared it as part of the Recovery Observatory project [7].

Finally, for the replicability demonstration over the 2023 June flooding event, raster data from the World Settlement Footprint (WSF) 2019 was accessed from the German Aerospace Center (DLR) [10]. This dataset consists of a 10 m resolution binary mask depicting the extent of human settlements worldwide and was derived from the 2019 Sentinel-1 and Sentinel-2 multi-temporal imagery, and thus it depicts the urban footprint at that year. It is worth noting that this is the latest version WSF product available for free from DLR Earth Observation Centre (EOC) geoservice, which offers services for visualising and downloading geospatial Earth observation data [11].

3.3. Generation of Amplitude Change Maps in Sentinel-1

In the second phase, amplitude change analysis was performed using SAR image pairs. Change maps as raster outputs were produced using ESA’s Geohazards Exploitation Platform (GEP). GEP is a thematic exploitation platform that provides online on-demand and systematic processing services for specific user community needs through advanced services available for both Optical and SAR data, connecting to massive computing power on multi-tenant Cloud Computing resources to address the challenges of monitoring tectonic areas globally [12,13]. GEP was exploited during the CEOS Recovery Observatory project as an infrastructure free of charge for the Haitian technical officers in order to support the mapping efforts during the recovery period, as well as for capacity building in remote sensing for future resilience to disasters [14].

In particular, two processing chains were used. The first is the COIN (Coherence and Intensity change) algorithm, which combines information from Sentinel-1 SAR images to detect changes in objects and on the ground. This algorithm uses SAR information as an intensity difference in the backscatter signal to detect changes in the amplitude of radar signals, which are related to the geometry and composition of the Earth’s surface [15]. The coherence was calculated between the two Sentinel-1 images in each pair, one as a master image and the other as the slave on different dates. The second algorithm is the SNAC amplitude change processing chain for Sentinel-1, which provides a SNAP-processed image pair, converting the intensity of the changes in the images into a result image [16].

Figure 6 shows an example of the results achieved during this step of the methodology. Specifically, this is a change map as a referred “ratio” between the backscatter in the pre-event SAR image and the backscatter in the post-event SAR image. The spatial extent of the ratio map matches with the overlapping between the pre-event and post-event SAR image frames.

3.4. Selection and Extraction of Relevant Change Data

In the third step, the relevant information highlighted in the change maps was extracted in GIS by applying a set threshold of SAR amplitude change measured in dB. The rationale and scope for thresholding are chosen to isolate significant changes that are due to the processes of interest from scattered and/or small changes that do not match with clear patterns, which are mostly due to random SAR amplitude oscillations and are not relevant for the analysis. SAR is a technique that is very sensitive to changes. Furthermore, SAR imaging is affected by speckle (i.e., noise); thus, it is common that applying amplitude change detection to SAR pairs, even if collected in a very short time span, may result in a wide distribution of changes within which only clear patterns are relevant. The threshold was then adjusted according to such sensitivity. In most cases the threshold of −5 and 5 was applied (i.e., to isolate changes in SAR backscatter that were 5 times more or less strong). In some specific change maps, it was necessary to increase or decrease this threshold to −6 and 6 or 4 and 4, respectively.

Table 3. Change detection maps with the specific threshold applied. Notation of the nomenclature of the change detection maps: sigmaDiff—backscatter ratio expressed in sigma nought; dB—decibel; IW—Sentinel-1 Interferometric Wide Swath Mode; VV—vertical–vertical polarisation; DDmonthYYYY indicates the dates of acquisition of the pre- and post- images, respectively. The standard threshold value was −5 and 5 (i.e., changes 5 times more or less intense). For some maps, it was necessary to set the threshold to −6 and 6 or −4 and 4 because these dates were affected by noise, so the threshold was adjusted according to this sensitivity.

Time Period	Image Pair Date	Range<	Range>
Pre-Event	sigmaDiff_dB_IW_VV_05Aug2016_29Aug2016.tif	−5	5
	sigmaDiff_dB_IW_VV_07Aug2016_31Aug2016.tif	−5	5
	sigmaDiff_dB_IW_VV_29Aug2016_22Sep2016.tif	−5	5
	sigmaDiff_dB_IW_VV_31Aug2016_24Sep2016.tif	−6	6
Cross-Event	sigmaDiff_dB_IW_VV_07Aug2016_18Oct2016.tif	−4	4
	sigmaDiff_dB_IW_VV_31Aug2016_18Oct2016.tif	−4	4
	sigmaDiff_dB_IW_VV_22Sep2016_10Oct2016.tif	−4	4
	sigmaDiff_dB_IW_VV_24Sep2016_18Oct2016.tif	−4	4
Post-Event	sigmaDiff_dB_IW_VV_10Oct2016_03Dec2016.tif	−5	5
	sigmaDiff_dB_IW_VV_18Oct2016_11Nov2016.tif	−5	5
	sigmaDiff_dB_IW_VV_11Nov2016_05Dec2016.tif	−5	5
	sigmaDiff_dB_IW_VV_03Dec2016_27Dec2016.tif	−5	5
	sigmaDiff_dB_IW_VV_05Dec2016_29Dec2016.tif	−5	5
	sigmaDiff_dB_IW_VV_27Dec2016_20Jan2017.tif	−5	5
Recovery Period	sigmaDiff_dB_IW_VV_20Jan2017_07Jul2017.tif	−6	6
	sigmaDiff_dB_IW_VV_22Jan2017_09Jul2017.tif	−5	5
	sigmaDiff_dB_IW_VV_07Jul2017_03Jan2018.tif	−5	5
	sigmaDiff_dB_IW_VV_09Jul2017_05Jan2018.tif	−5	5
	sigmaDiff_dB_IW_VV_03Jan2018_19Aug2018.tif	−5	5
	sigmaDiff_dB_IW_VV_05Jan2018_21Aug2018.tif	−5	5
	sigmaDiff_dB_IW_VV_19Aug2018_10Jan2019.tif	−5	5
	sigmaDiff_dB_IW_VV_21Aug2018_24Jan2019.tif	−5	5
	sigmaDiff_dB_IW_VV_10Jan2019_14Aug2019.tif	−5	5
	sigmaDiff_dB_IW_VV_24Jan2019_16Aug2019.tif	−5	5
	sigmaDiff_dB_IW_VV_14Aug2019_05Jan2020.tif	−5	5
	sigmaDiff_dB_IW_VV_16Aug2019_07Jan2020.tif	−5	5
	sigmaDiff_dB_IW_VV_05Jan2020_01Sep2020.tif	−5	5
	sigmaDiff_dB_IW_VV_07Jan2020_03Sep2020.tif	−5	5

lists all the image pairs analysed, with indication of the threshold applied.

Thresholding led to the selection of SAR amplitude change pixels that were converted to vector format to facilitate further geospatial analysis of the found changes. Given that the natural hazard subject to analysis was a hurricane that induced destruction and flooding, the main targets of interpretation in the found changes were housing, water bodies and floods that occurred as a result of the phenomenon.

In order to identify whether the found changes were related to damaged, destroyed or reconstructed houses and dwellings, we exploited the shapefile of humanitarian OSM [17] as the geospatial reference of all the dwellings existing prior to the hurricane occurrence. Further means for verification were also provided by the EMSN050 dataset, which classified the affected buildings in the two sectors of Jérémie and Les Cayes (see Figure 3 for spatial location and extent of these sectors). Figure 7a shows an example of the geospatial intersection between the OSM dwelling inventory and the extracted SAR amplitude change pixels. This intersection allowed us to associate the found changes to the dwellings.

From a methodological point of view, when the outcome of the above geospatial intersection led to an apparently high number of affected dwellings, thus suggesting a potential overestimation of damages, a manual check was made by visually inspecting the pre- and post-event images. Although Sentinel-1 images are provided as normalised backscatter products, it may happen that one image may exhibit a significantly and generally higher backscatter signal, which the ratio operation may emphasises.

The above geospatial intersection and the following comparison with the visual evidence provided by temporally co-located optical satellite image(s) also helped to cope with another possible source of overestimation. Given the coarse spatial resolution of Sentinel-1 and the mapping scale, it could happen that a pixel of the SAR images containing a single dwelling could exhibit changes even if the dwelling was not damaged or changed. This could be due to the effect of SAR amplitude change in the surrounding trees, vegetation or other types of elements.

Finally, particular attention was paid, especially for the analysis in the emergency period (i.e., soon after the event and for the months October–December 2016), to the appearance of SAR amplitude change pixels exhibiting a clear and strong increase in SAR backscatter. These changes are clear proxies of emergency and post-emergency activities such as installation of temporary constructions like humanitarian aid camps. An example is shown in Figure 7b. Tracking the persistence and later the disappearance of these signals allowed the identification of the installation and dismantling phases. Of course, because these structures were not present prior to the event, their detection was based on the evidence that they fell in areas where no buildings were inventoried in the ancillary data, e.g., in the OSM dataset.

For flooded areas, the detection was made against permanent pre-existing water bodies. The intersection with the land cover dataset and knowledge of the permanent water bodies and rivers were crucial. SAR amplitude change pixels related to inundated areas were selected on almost all dates by applying a threshold between −5 and 5. In this way, flooded areas and the greatest impacts on mangroves were mapped. An example is reported in Figure 7c.

3.5. Analysis of the Amplitude Changes

Once the data with backscatter change values between corresponding dates were obtained, an analysis of the acquired information was conducted. For dwellings structures, the number of pixels that experienced changes was quantified to comprehend the temporal dynamics of changes in dwellings during the immediate aftermath of the hurricane and the recovery period. The change patterns were proxies for the following situations: no damage, damaged, destroyed/disappeared, reconstructed and new construction.

Regarding flooded areas, changes between pixels were analysed and quantified to understand how they expanded or contracted based on the fluvial and geomorphological dynamics of the area. Positive amplitude change values indicated areas that had flooded, since the backscatter value was higher in the first image (pre-event) compared to the second (post-event). Conversely, pixels with low or decreased backscatter values typically corresponded to specular reflection from water bodies, thus leading to lower backscatter values. When flooding was detected in change maps produced from earlier SAR image pairs (i.e., those closer or precisely falling in the period of the hurricane), the negative amplitude change values found in the later image pairs (thus temporally distant from the hurricane) were used as proxies of water recession. The backscatter in the second image (more recent) was greater than the first image (older) due to the higher backscatter from exposed dry rough surfaces or vegetation (see the example in Troupeau in Figure 8a).

3.6. Generation of Maps of the Hurricane Impacts and Recovery in the Island

The final output of the proposed methodology is a set of thematic maps showing the impact due to the hurricane and the progress in recovery and reconstruction. In order to obtain maps of the hurricane impacts that provided detailed information on the extent and areas before and after the event, maps were generated for each interval using the ascending and descending geometry image pairs (see the example in Jérémie in Figure 8b), and spatial information of the changes observed during the different time phases studied, before and after the hurricane.

A detailed analysis of the information from the areas most affected by the hurricane was carried out in order to determine the extent of the damage caused by the event and the possible long-term consequences in the study area. During this analysis, we could observe the extent of the hurricane’s impact on homes and areas with water bodies and areas susceptible to flooding. Areas with significant changes in dwellings, flooding areas and changes in mangrove vegetation were identified. In addition, this analysis made it possible to assess the likely evolution of these areas in the future, considering the long-term consequences of the hurricane. For example, mapping areas where the risk of flooding is present and is verified after a real hurricane event is the evidence base for the planning of preventive measures to be taken to reduce possible negative impacts in future.

4. Results

4.1. 2016 Post-Event Assessment

4.1.1. Flooded Areas

Figure 9 shows the two maps of flooded areas as derived from the amplitude change detection of the cross-event Sentinel-1 image pairs collected along both the ascending and descending orbits between 22 September and 18 October 2016. These thematic maps are multi-temporal, given that they combine all the changes due to flooding and the following water recession as they were captured by the various images composing the Sentinel-1 time series. The blue colour indicates flooded areas, whereas the red colour indicates the land portions that have dried up during the cross-event.

Three main areas are found. The main flooded areas are located in the region east of the town of Troupeau. The second most affected region was the area around the Miragoane lagoon, located far east of the study area. In addition, there is an area with several permanent lakes that reach their maximum hydrological level during the rainy season, north of the agricultural district of Les Cayes. The other areas impacted by the floods are the lowlands of the main rivers that flow through the study area, like the river Grande-Anse that flows into the Caribbean Sea near the town of Jérémie. Figure 10 shows the matching evidence of the riverine flooding and sediment plumes towards the sea found in the Sentinel-2 optical images.

The operational advantage of such thematic flood extent maps is twofold: (1) the maps provide an at-a-glance spatial assessment of the hurricane-induced flooding; and (2) they allow the operator to quantify the spatial extent and the temporal persistence of flooded areas at each location.

Figure 11 details the three main areas affected after the passage of Hurricane Matthew. Prior to the event, ephemeral and random changes occurred mainly in the Troupeau region, where local surface geology and hydrology cause conditions for surface water accumulation and drying out (Figure 11a). In the Miragoane region during the pre-event, changes were observed, with a decrease in the radar signal within the lagoon, but these were mainly due to the wind-induced change in the wave lake structure (Figure 11b). For the lakes north of Les Cayes, the increase in the radar signal was associated with a drying-up phase. The smaller water bodies were completely dry, while concentric change patterns along the border of the larger bodies clearly highlight the natural process of water recession (Figure 11c). Subsequently, during the cross-event, the situation on the ground changes completely. In Troupeau two major change patterns were found: one to the north and the other one to the south (Figure 11d). According to the Sentinel-1 results, the southern part dried up more quickly than the northern part of Troupeau (Figure 11d), with the latter taking up to December to completely dry out (Figure 11g). This spatio-temporal assessment is helpful to estimate when the land could be recovered to rejuvenate the cultivations in this area, which is essentially a rural district where farming activities are undertaken by the local population.

From a quantitative point of view (Figure 12), the hurricane caused a total increase in flooded areas extending across land of more than 20 square kilometres. At the same time, the curve related to the decrease in flooded areas highlights a linear ramp up towards November 2016. The decrease is mostly concentrated over the areas that were affected by the hurricane’s passage and testifies to the gradual natural recovery in the absence of further rainfalls and storms. As found in Troupeau (Figure 11d,g), as well as in the lakes north of Les Cayes (Figure 11f,i), it took up to December 2016, i.e., two full months after the hurricane’s impact in December, for the majority of the flooded areas to dry up (Figure 12).

4.1.2. Affected Urban Areas

Figure 13 shows how impactful the hurricane was on the urban environments and assets of both the towns of Jérémie and Les Cayes. The maps are obtained by combining the change patterns detected from both the ascending and descending geometries and are thresholded as indicated in Section 3.4. The greater increase in backscatter is associated with the accumulation of debris and roofless structures, while the decrease in backscatter reflects the destruction of the buildings, thus the structures acting as radar reflectors. During the post-event period up to the end of 2016, there was no significant change or sudden increase in the signal return, meaning that very few buildings were rebuilt, and the rubble and waste were collected. This observation matches with the information from the ground, highlighting the dramatic situation for the weeks and months after the hurricane, as well as the spread of temporary shelters and humanitarian aid camps.

In addition to the main urban centres, the hurricane destroyed several dwellings also in the rural areas of central Haiti. Figure 14 shows a clear example in the central mountainous range, where the change patterns found in Sentinel-1 image pairs spatially match the areas of land and topography that were exposed to the hurricane winds coming from the south. The change pattern concentration reveals a higher damage compared to the leeward slopes or behind the mountain.

Figure 15 shows the quantitative analysis of the change patterns related to building damage. The graphs of both increase and decrease in backscatter exhibit the same behaviour, with one peak temporally centred during the cross-event interval. This analysis confirms that the observed changes are all related to hurricane destruction and that no real evidence of reconstruction is detectable at the mapping scale allowed by Sentinel-1 in the months soon after the event.

4.2. Analysis of Results for the Recovery Period from 2017 to 2020

The recovery period that was investigated included years from 2017 to 2020 in order to document the mid- to longer-term pathway of the Haitian community in recovering from Hurricane Matthew impacts. This period also matches with the monitoring period of the CEOS Recovery Observatory and relates to local stakeholder needs. For each year, two 6-month observations and associated assessments were made. The key result was the evidence of a greater acceleration in reconstruction in some areas, while in others the reconstruction process was less effective and slower.

4.2.1. Results in Flooded Areas

During the recovery period, the areas flooded by the hurricane did not show large changes in backscatter differences. This is effectively shown by the multi-temporal thematic maps displayed in Figure 16 over the Troupeau area. Not surprisingly, the greatest changes were found during the two semesters of the year 2017, with a steady decrease in the backscatter or drying of these flooded areas. This result also confirms that, in the absence of a major weather event driving inundation in Haiti, there are no local sources of flooding that could cause persistently flooded areas, which may happen instead if a hurricane or a storm induces the failure of dams or riverbanks and thus the occurrence of a dam/riverine inundation. In such circumstances, the persistence of flooded areas may be observed for months, or even years (see the example in Colombia investigated by Velasquez Hurtado et al. [18]).

The quantitative time-series analysis of flooded areas combining the recovery period and the 2016 post-event assessment (Figure 17) clearly outlines that changes related to flooding and water recession were constantly below an area of 6 km² during the recovery phase. The time series highlights that no extreme events or natural disasters occurred in the region.

4.2.2. Analysis of Results in Built-Up Areas

Within the eight sectors that were analysed in detail, the results show greater changes in the two largest towns in the study area, i.e., Jérémie and Les Cayes. In Jérémie (Figure 18), greater or noticeable changes were observed according to the change map during the year 2017. Decreased backscatter clusters were found across the whole urban footprint, suggesting that 2017 was characterised by the main changes associated with reconstruction, starting from the removal of rubble and demolition of damaged buildings. In particular, the second semester of 2017 recorded the highest increase in changed area after the emergency period (see quantitative information displayed later Figure 21). Similar patterns but different signs are found in 2018, with clusters of increased backscatter highlighting where new and restored buildings and dwellings were located. This suggests that reconstruction mostly happened by that time. At the same time, the comparison with cloud-free Sentinel-2 images shows that, from 2017 to 2019, some constructions disappeared, and in most of the cases they were temporary shelters and emergency assets that had been installed during the emergency phase and were later dismantled (Figure 19a). Another phenomenon that was observed is related to the structures being built near the coast, and even along the coast (Figure 19b). This is frequently associated with spontaneous settlement in areas that were strongly affected by the sea waves and wind during the hurricane, and thus should not be populated.

Similar patterns and temporal evolution were found in Les Cayes (Figure 20), although here the greatest increase in the number of new constructions was observed (Figure 21). The semester with the most significant change data was the first half of 2017, although, towards the end of the recovery phase in 2020, new increases in the results of the change maps emerged (Figure 20).

4.2.3. Results of the Sentinel-2 Optical Image Classification

Sentinel-1 backscatter changes were used as a proxy of new building construction or, in the opposite case, for demolition and dismantlement. To further strengthened this interpretation, Sentinel-2 optical images were used to generate urban land cover classification maps. Figure 22 and Figure 23 show the results for the two main sectors, i.e., Jérémie and Les Cayes. The maps indicate that the semesters with the highest number of new constructions were the second semester of 2017, followed by the last two semesters analysed in this study (II 2019 and I 2020).

In Jérémie, the largest area of new construction was located in the rural hilly land north-west of the city (Figure 22). Figure 22a shows an example of land cover transition from rural to built-up due to new construction. Another relevant class of changes is related to new infrastructure. Figure 22b shows an example where a newly built road and new buildings can be visually detected. Figure 22c highlights how another sector of the town that significantly changed was the coast south-east of Jérémie, where the Grande-Anse River flows. As also found by De Giorgi et al. (2021) [19] using COSMO-SkyMed SAR and Pléiades optical images, this stretch of the coast was affected by the growth of spontaneous settlements and mining and quarrying activities, especially by the poorest local population during the years following the hurricane.

In Les Cayes, the changes due to new constructions were found at later dates, mainly to the north of the town and west of the urban centre, along the road network (Figure 23).

4.3. Replicability to 2023 Flood Event

As detailed in Section 3.1.2, a flooding event happened on 2–3 June 2023. The change map was produced from the Sentinel-1 cross-event image pair 27 May–8 June 2023 to identify the most affected areas and/or exhibiting the greatest number of changes in backscatter after thresholding with a value equal to 5 (Figure 24a). To select the urban areas, spatial intersection with the WSF was carried out, and the output shapefile was exploited to calculate the extent of the damage (Figure 24b). The EMSN050 administrative boundary vector layer was used as a reference to present the results by districts.

Figure 25 shows the three main areas where inundation was found, thus outlining at a glance that the 2023 June event affected many areas across the country. Not surprisingly, as in October 2016, the Troupeau area was flooded (Figure 25a–c). The 27 May event indicates that the plain was dry, with no evidence of a surface water body (Figure 25a). After the event, several spots show surface water accumulation. Although the severity was much lower than the 2016 event (Figure 25b), it is interesting to note that the flooded areas were located within the same boundary of the 2016 inundation (Figure 25c, to be compared with Figure 11d).

The other areas affected by flooding were, to the north, the urban areas in Desdunes (Figure 25d) and the Leogane area, west of the capital Port-au-Prince, where large agricultural areas were inundated (Figure 25e). From a quantitative point of view, the Fossé Naboth district, north of Port-au-Prince, was the district with the highest impact and the most flooded areas (

Table 4. Statistics of the 2023 June event impacts; spatial extent in square metres of the affected areas in each district/town.

Region	Town	Square Metres
	Bas Coursin	16,393.9
	Bélanger	4693.2
North of Port-au-Prince	Desdunes	11,599.0
	Fossé Naboth	88,381.7
	Labady	1073.8
	Villars	34,148.0
Region 1	Total	156,289.6
	Dessources	527.6
West of Port-au-Prince	Grande Rivière	6359.0
	Gros Morne	2668.6
	Total	9555.2

¹ Denotes the other region.

and Figure 26).

5. Discussion

From an implementation point of view, the proposed methodology proves to be effective in detecting changes in the different phases (emergency, aftermath, reconstruction and recovery) following the occurrence of hydrometeorological hazard events such as hurricanes, tropical storms and flooding. The reliance on SAR backscattering for change detection has the advantage of a high sensitivity to surface changes, in addition to operability during the cloudy weather that is typical of both such events and the tropical climate zone. Therefore, it is very unlikely that relevant changes are missed, and even changes at the single pixel level are successfully identified. The latter capability is especially enhanced in predominantly rural environments such as the one characterising Haiti, so that even isolated new dwellings can be spotted, despite the coarse resolution provided by Sentinel-1 data. However, the drawback is that ephemeral and unrelated changes may also be detected. Therefore, filtering is needed, e.g., via SAR backscatter amplitude thresholding. It is recommended that the latter step should be implemented by adapting to the given environmental and anthropogenic context and in relation to the value distribution before thresholding. This operation implies an operator-driven approach and is applicable to change maps generated from image pairs covering both emergency and recovery phases.

In general, as expected, the achieved results indicate a greater change in the values immediately after the passage of the hurricane during the “cross-event”. A careful investigation of the spatial distribution of the affected areas does not only provide a correlation with the hurricane pathway, but also highlights where recovery and reconstruction should mostly happen, and thus be seen from satellite data in the later months/years. Therefore, the thematic maps can serve as a geospatial means of identifying hotspot areas to monitor over time. For example, during the analysis, it emerged that houses or structures located in the valley areas between the hills or in the mountains, which offered some protection from the hurricane’s strong winds, were not as severely affected as those directly exposed to the hurricane’s winds on the windward slopes. This differentiation helps to select which areas must be monitored. Another example occurred along the rivers and their mouths. The hurricane led to a massive increase in the flow of most rivers, which was detected through the change maps. As a result, downstream, the morphology of the river mouths along the coastline was altered, and the existing infrastructure and dwellings in the nearby area were found to be impacted. Those areas must be considered high-risk if new construction (either legal or spontaneous) happens and the population relocates there; they should be designated as high-risk zones, and planning regulations should discourage new construction or population relocation in these vulnerable locations.

In the case of Haiti, once the first months passed since Hurricane Matthew, in 2017, there was another boost (albeit much smaller in magnitude) in surface changes that highlighted the initial recovery and reconstruction actions. The spatial and temporal mapping, alongside the quantitative assessment, suggested that recovery and reconstruction mostly happened in 2017 and 2018 in both Jérémie and Les Cayes. Therefore, the proposed methodology was effective in assessing if the process was ongoing and locating the main hotspots. It was then, at that stage, also possible to infer potential weaknesses in the recovery and reconstruction process. This is exemplified by the evidence found along the coastline in front of the Grande-Anse river mouth, south of Jérémie. Sentinel-1-based thematic maps and the visual inspection of temporally co-located high resolution Sentinel-2 images were sufficient to show these settlement dynamics and anthropogenic activities, providing evidence comparable with similar observations made based on much higher resolution satellite imagery (see comparison with De Giorgi et al., 2021) [19].

The replicability test, undertaken by applying the proposed methodology on the event that happened in 2023, proves that the workflow is effective for achieving a country-scale post-event assessment and, at the same time, measuring quantitatively the impact at a district level. This test was limited to the emergency phase and did not include a multi-year recovery monitoring because the scale and severity of the event and its impacts were lower than the 2016 event. However, the achieved results outline how the method is replicable and can be carried out even with limited computing resources and operators.

In this respect, while the methodology relies on simple and consolidated processing routines, and although many more sophisticated change detection approaches exist, it is clear that one of its key strengths relates to the absence of real technical barriers for its implementation by non-specialist users and beginners. Although its workflow can also be performed by inputting proprietary/licenced and commercial SAR and optical data, it is independent from the availability of this type of data. Furthermore, the workflow can include additional data that, in the present configuration, are not encompassed but may be advantageous for retrieving key physical parameters useful for assessing hydrological and weather-related hazard events, such as LiDAR data. These data provide crucial topographic information that the satellite imagery exploited in this demonstration case study cannot provide in terms of spatial and temporal resolution. However, LiDAR data is often difficult to access for decision-makers, like risk planners and emergency responders, due to the costs and challenges associated with data acquisition and processing. Therefore, the addition of LiDAR data, as well as any other data, should account for feasibility and sustainability so as not to compromise the workflow usability by users.

Another aspect that must be acknowledged is that the mapping detail and scale are a function of the input data. The main limitations to account for are as follows. First of all, the high SAR sensitivity to surface changes may lead to overestimation and noise, and thus expert knowledge is advantageous and, additionally, interpretation by the operator is required. Prior knowledge of the local context expedites the selection of the relevant changes and their interpretation, and can inform filtering out of the unrelated changes. This is particularly important when the input data are like that of Sentinel-1, and verification with optical data may be limited given the lack of cloud-free imagery. Therefore, one approach to improve the method that future research should investigate would be to test how machine learning could help in pattern recognition and extraction, and subsequent classification, thus increasing the automation and objectivity of the whole workflow.

6. Conclusions

When a disaster affects a territory, most of the attention is focused on the emergency phase. However, Earth observation data and GIS can help in supporting monitoring efforts in the months and years after to assess if the affected territory is recovering. This research aimed to demonstrate that the latter task can be addressed by means of a feasible and cost-effective methodology relying on free and open satellite data, as well as simple and consolidated processing routines and geospatial analysis. Indeed, there are several methodologies for exploring and analysing optical and SAR data to achieve this purpose. However, a key question is who will be able to implement it, especially in low-income countries. Therefore, it is important to build and demonstrate the usefulness of methodologies to be used for long-term comparisons and analysis of the recovery process that are reliable and, at the same time, usable without technical barriers.

The present study has provided a real-world scenario demonstration to assess the impact of Hurricane Matthew in the south-western departments of the island of Haiti and evaluate the spatial distribution and temporal evolution of the reconstruction and recovery activities over the four years after the event. The multi-temporal thematic maps produced by the change detection of Sentinel-1 image pairs and the verification with temporally co-located cloud-free Sentinel-2 images and free access Planet imagery led to (i) the detection of the most affected areas and quantification of the extent of damaged buildings and inundations; (ii) the measurement of how long it took for flooded areas to dry up and waters to evacuate; (iii) the locating of temporary shelters and emergency assets and their dismantlement; (iv) the identification of areas where reconstruction and new building construction happened and when this happened; and (v) the identification of areas where reconstruction and/or recovery activities occurred in susceptible areas, and thus to the detriment of future resilience to similar disaster events. This analysis was undertaken at country scale and thus was helpful for identifying specific locations as candidates for conducting further detailed monitoring studies. Although fine detail mapping was prevented by the coarse resolution of the majority of the input satellite data, the methodology allowed us to document the impact and assess the recovery in many locations to which access after the hurricane was hampered due to storm debris or road closures, and remained hampered for months afterwards, or that were unfeasible to inspect in situ.

The thematic maps and associated statistics retrieved for the main towns of Jérémie and Les Cayes prove effective for supporting decision-makers to (i) determine priorities in response and recovery planning, land-use planning and risk mitigation measures, (ii) track the progress of these actions and (iii) identify any deviations, such as new settlement patterns in at-risk areas.

While the methodology proves to be replicable in addressing mapping and monitoring purposes in similar events, it shows the disadvantage intrinsic to the coarse resolution of the input data, i.e., Sentinel-1 images. As a consequence, sometimes invalid data or outliers may occur and, given the specific land cover and spatial arrangement and proximity of urban features and vegetation, the backscatter contributions at the pixel level may be mixed and thus lead to an overestimation of surface changes. Future improvements should be sought in applying the methodology to high-resolution data where available and affordable, such as TerraSAR-X and COSMO-SkyMed SAR datasets, and testing machine learning approaches to automate pattern detection, extraction and classification. However, if these improvements are not achieved in a way such that the methodology remains accessible by non-specialists and those operators lacking resources, the cost-effectiveness and usability of the methodology will be significantly hampered.