Deciphering climate-induced displacement in Somalia: A remote sensing perspective

RESEARCH ARTICLE Deciphering climate-induced displacement in Somalia: A remote sensing perspective Rahman Momeni1, Tuba BircanID2*, Robert King1, Eloy Zafra Santos1 1 GMV, Harwell, Oxfordshire, United Kingdom, 2 Dept. of Sociology, Vrije Universiteit Brussel, Brussels, Belgium * tuba.bircan@vub.be Abstract Rapid climate changes bear significant consequences on various aspects of our lives, nota- bly by deteriorating living conditions in certain areas to such extent that inhabitants have no choice but flee. Despite recognition of this issue, the dynamics of the relationship between the environmental factors and the human mobility have yet to be thoroughly investigated. This study aims to explore the application of advanced remote sensing analytics for devel- oping detailed climate indicators at a micro (district) level, and to examine the relationship between climate factors and internally displaced persons. After detailing our data sources and the analytics employed for indicator development, we discuss various types of events and their repercussions. Our findings corroborate that slow-onset and rapid-onset climate events differently impact society, and the responses hinge on the urgency precipitated by the detrimental aftermath of the extreme weather event and, most crucially, on people’s capabilities. We also underscore the importance of data quality and availability for the socio- economic indicators to enhance future studies, given the intertwined associations between climate change, economic deprivation, and violent conflict. Introduction

Over the past decade, we have witnessed numerous shocks across the globe - whether political, economic, societal, or environmental - triggering humanitarian crises and prompting millions to flee or be displaced. Political instability in the Middle East and economic downturns in Latin America since 2015, for instance, have dominated international discourse as mass dis- placement ensued, with individuals initially seeking refuge in neighbouring countries before looking further afield [1–5]. Such crises, and the resultant emigration they provoke, have been extensively documented and rigorously studied. As of February 2022, global attention has turned towards Ukraine as it confronts its own displacement struggle. Whilst these sudden - albeit not unexpected - events typically attract significant attention due to the immediate response required from host countries, numerous nations, particularly in Asia and Africa, have been contending with the severe impacts of chronic or acute environmental disasters for years. When compounded with enduring chal- lenges like poverty, famine, armed violence, and political instability, these environmental PLOS ONE PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 1 / 27 a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Momeni R, Bircan T, King R, Santos EZ (2024) Deciphering climate-induced displacement in Somalia: A remote sensing perspective. PLoS ONE 19(8): e0304202. https://doi.org/10.1371/ journal.pone.0304202 Editor: Bijeesh Kozhikkodan Veettil, Van Lang

University: Truong Dai hoc Van Lang, VIET NAM Received: May 23, 2023 Accepted: May 8, 2024 Published: August 9, 2024 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal.pone.0304202 Copyright: © 2024 Momeni et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All data are publicly available. To be more specific, - The aggregated and anonymised tabular data used for this paper is available under the link: https://github.com/ bircantuba/HMB.git - The OSF data (for the remote

factors’ societal effects amplify, forcing people to relocate. As a result, these individuals first experience displacement within their own countries before eventually seeking refuge elsewhere. Indeed, the global population of internally displaced persons (IDPs) has steadily increased, with the 2021 Internal Displacement Index Report indicating that by the end of 2020, the num- ber of IDPs worldwide had surpassed 55 million, double the number of refugees [6]. By devel- oping a nuanced understanding of the root causes behind irregular migration and forced displacement, we can better anticipate patterns of human mobility. More importantly, we can enhance protection measures for migrants and refugees, providing them with the safety and support they need in these challenging circumstances. The field of climate-induced migration has seen significant advances in contemporary pre- dictions compared to earlier efforts, spurred by growing interest in studying environmental dynamics and mobility decisions [7–12]. Harnessing innovative data sources and cutting-edge analytics has enormous potential for data collection, consequently enhancing our understand- ing of human mobility [13, 14]. This improved knowledge can be instrumental in the develop- ment of predictive models [15, 16], understanding slow-onset processes of climate change [17], and advancing socio-economic indicator development [18]. Particularly, remote sensing technologies offer unique capacities to study the relationship between the climate change and human mobility. These technologies provide critical environ-

mental data at varying scales, elucidating the impacts of climatic indicators, such as droughts and floods, on population movements [19]. Satellite-derived data offer insights into climatic trends, including rainfall variability and temperature anomalies, alongside visualising land use changes and environmental degradation [20, 21]. The integration of these datasets with socio- economic variables can reveal patterns and hotspots of climate-induced migrations [22]. Moreover, remote sensing data, coupled with machine learning algorithms, can be used to pre- dict future migration patterns under various climate scenarios [23]. With this study, we will assess how to maximise the potential of these technologies in understanding the intricate inter- play of climate change and human displacement. Accordingly, the primary objective of this paper is to scrutinise the relationship between cli- matic factors and IDPs within the context of Somalia from 2016 to 2019. Our approach employs advanced remote sensing analytics to construct sophisticated climate indicators at a granular, district level. The structure of this article unfolds as follows: We begin with a comprehensive examination of the case of Somalia, elucidating both the environmental conditions and the historical narra- tive of migration. Subsequently, we delineate our data sources and the methodological frame- work used for creating climate indicators. In the section analysing our results, we discuss our investigation and visualise the indicators through maps to elucidate the relationship between

environmental factors and IDPs. In the conclusion, we summarise our findings and offer insights and recommendations aimed at enhancing current practices in this research domain. Environmental change and displacement in Somalia Our study focuses on Somalia due to its vulnerability to extreme climatic events, which have triggered considerable internal migration. The nation’s fragility and susceptibility stem from a combination of enduring conflicts, protracted epidemic outbreaks, erratic climatic conditions and shocks adversely affecting agriculture, livestock, and rural livelihoods, which are funda- mental for food security. Simultaneously, weak social protection systems further exacerbate the situation. The World Bank classifies Somalia as one of the world’s most impoverished countries, with approximately 70% of its population living below the poverty line, subsisting PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 2 / 27 sensing analysis) can be found under the link: https://osf.io/kg9zd/ Funding: MR, BT, KR This research is supported by the European Commission through the Horizon2020 European project: “HumMingBird – Enhanced migration measures from a multidimensional perspective” (GA: 870661). https://research-and-innovation.ec.europa.eu/ funding/funding-opportunities/funding- programmes-and-open-calls/horizon-europe_en The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

on less than $2 per day. Furthermore, the United Nations High Commissioner for Refugees (UNHCR) places Somalia seventh globally in terms of displaced populations. Since 1990, Somalia has endured over 30 extreme weather events, encompassing 12 droughts and 19 floods, triggering two significant famines in 1991–92 and 2011. These disas- ters led to more than 3.2 million new displacements, implying the number of movements rather than the number of individuals displaced. Given that over 80% of Somalia’s territory features arid and semi-arid climates, and rainfall exhibits high spatial and temporal variability, these adverse environmental impacts and inadequate natural resource management negatively affect the nation’s economy. Approximately 70% of Somalis are reliant on climate-sensitive agricultural and pastoral activities. The recurring interplay of drought and floods undermines food security and access to basic services, heightening vulnerability and prompting displacement at varying scales. UNHCR data reveals that since 2016, over three million Somalis have been displaced due to recurrent extreme droughts and floods. Despite assumptions concerning "environmental migration" that posit immediate displacement following weather-related disasters such as floods, storms, and drought, the displacement process involves diverse responses and trajectories influenced by contextual conditions and individual capabilities rather than purely rational choices. In regions like Somalia, prone to recurrent extreme weather events, disasters and detrimental

environmental conditions may trigger multiple instances of displacement, further amplifying vulnerability. Environmental factors, progressively worsening as the climate becomes hotter and drier, directly impinge upon crop, livestock, and forestry-based livelihood systems. Regular and pre- dictable rainfall is critical for the pastoral and rural communities across Somalia. Beyond the immediate effects of abrupt extreme weather events, the slow-onset of environmental degrada- tion is a crucial factor provoking displacement towards more habitable regions. These rapid- and slow-onset environmental conditions may play divergent roles in individuals’ decisions to migrate either internally or internationally at different times. Numerous qualitative and mixed-methods studies have shown that in many instances, the decision to migrate often materialises as a "last resort" in response to an extreme event, rendering existing living condi- tions unbearable [24–26]. To distinguish between new displacements and IDPs, Fig 1 denotes the number of movements, not individuals, as individuals can be displaced multiple times, a Fig 1. Total number of new displacements due to disasters between 2008–2020 in Somalia. https://doi.org/10.1371/journal.pone.0304202.g001 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 3 / 27

factor not always reflected in data collected by the Internal Displacement Monitoring Centre (IDMC). Fig 2A presents the population density across Somalia, using a granular 100 × 100-meter grid square. Detailed views are shown around (a) Hargeisa, (b) Bosaso, and (c) Mogadishu. In Fig 2B, shows the distribution of two primary rivers in Somalia refers as breadbasket of Soma- lia with irrigated agriculture land producing Somalia food production. Shebelle, spanning the central areas, and Juba, crossing the southern regions of the country. The land cover of Soma- lia, as depicted by the European Space Agency (ESA) under the Africa land cover project in 2016, is represented in Fig 2C. Lastly, Fig 2D illustrates the proportion of population of each region displaced from their homes to other regions due to environmental factors. The IDP data was acquired from United Nations Human Rights Council (UNHRC). This map gener- ated by aggregating the percentage of the population of each region displaced due to flood and drought between 2016 and 2019. These aggregate figures provide a macroscopic overview of Somalia’s severe and erratic climatic conditions and the resultant human responses, often marked by difficult decisions to abandon their livelihoods. However, these annual figures inad- equately elucidate the rapidly changing yet gradually evolving environmental dynamics and their indirect and direct impacts on human lives. Migration decisions can manifest both as an immediate reaction to disasters or a longer-term choice stemming from recurring environ-

mental degradation and bleak long-term prospects. Given the diversity in individuals’ responses to similar circumstances, dictated by their physical and economic limitations and their perception of short to long-term risks, it is challenging to discern the degree to which these movements contribute to the continuum of displacement. Thus, more time-sensitive environmental indicators can provide valuable insights into the relationship between climate alterations and internal displacement. The scope of our study, delineated by the bold-blue line in Fig 2D, encompasses regions sig- nificantly affected by adverse climatic conditions, resulting in internal displacement. Criteria Fig 2. Maps of Somalia. a. The population density map across Somalia; b. Map of the major rivers in Somalia; c. Land- cover map of Somalia; d. The proportion of the population displaced across Somalia at regional level between 2016 and 2019. The extent coordinates for Somalia: WGS 84 (EPSG:4326) are 41, -1.7 to 51.4, 12. https://doi.org/10.1371/journal.pone.0304202.g002 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 4 / 27

for area selection involved considerations of population density, proximity to major rivers, and the degree of reported internal displacement attributable to environmental factors. Central and Southern Somalia were selected for this study, as these regions are not only densely popu- lated (Fig 2A) but are also bisected by the Shebelle and Juba rivers (Fig 2B). This geographic configuration provides conducive conditions for crop cultivation and forage production (Fig 2C). Furthermore, these regions have consistently confronted internal displacement due to extreme environmental events (Fig 2D). Fig 3 delineates the study area at the district level, with a total of 16 districts selected from the Central and Southern parts of Somalia, covering an area of 111,578 km2. Detailed descrip- tions of the selected districts are provided in Table 1. Data Our objective is to scrutinise the patterns of environmentally-induced internal displacement in Somalia by analysing the indicators we derived, aiming to discern any relationships with IDPs throughout the Area of Interest (AOI). We aim to understand the connection between extreme weather events, slow-onset climate change, and internal displacement in Somalia between 2016 and 2019. This study spans from January 1, 2016, to December 31, 2019. We chose the start year 2016 due to the availability of data from the Sentinel-1 (the first satellite in the Copernicus Programme launched by the European Space Agency) and Sentinel-2, an Earth observation mission from the Copernicus Programme that systematically acquires opti-

cal imagery at high spatial resolution over land and coastal waters, which we used to generate flood and drought maps. Furthermore, migration data has been available since 2015. Previous studies that take a quantitative empirical approach to the climate-migration rela- tion can be divided into micro studies, which largely focus on individual and household migra- tion [26], and macro studies, where researchers examine environmentally induced migration at regional or national levels [9]. Regarding different hazard types, many studies focus on fac- tors associated with climate change, such as changes in precipitation and temperature levels and variability [12]. This paper represents one of the first attempts to analyse climate/environ- ment-induced displacement at the district level, employing environmental indicators derived from daily, weekly, and monthly remote sensing/satellite data. Fig 3. Study area consist of sixteen districts with 111,578 km^2 coverage throughout centre and South Somalia. https://doi.org/10.1371/journal.pone.0304202.g003 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 5 / 27

Remote sensing data Sentinel-1. Sentinel-1 offers an open-source Synthetic Aperture Radar (SAR) imaging sat- ellite data with medium spatial resolution as part of the European Space Agency Copernicus programme. Sentinel-1 comprises of two satellites launched on 3 April 2014 and 22 April 2016. The SAR system operates within C-band (5.407 GHz) frequencies in one of four acquisi- tion modes, with the default mode being Interferometric Wide Swath on land. In total, we downloaded and processed 32 SENTINEL-1 GRD scenes (16 pairs) from the Alaska Satellite Facility (ASF), spanning from 07/01/2016 to 30/10/2019. We used the Sentinel-1 satellite data to map the flood events in Somalia. Fig 4 shows Sentinel-1 acquired on 30th May 2016 during a flood event affected Beledweyne city. The areas near riverbank with black colour represent as the flooded zones where there are low backscatter. Furthermore, it has been observed that sandy areas located further from the river exhibit a similar low backscatter to the flooded areas and also represented in black. Sentinel-2. The Sentinel-2 (Sentinel-2) dataset is an open-source optical with medium to high spatial resolution satellite data provided as a part of the ESA Copernicus programme. The S-2 mission consists of two satellites: Sentinel-2A and Sentinel-2B. Sentinel-2 provides multi- spectral data with 13 bands in the visible, near-infrared, and short-wave infrared parts of the spectrum. The S-2 is equipped with the state-of-the-art Multispectral Instrument, offering

Table 1. List of districts used as areas of interest (AOI) for agricultural drought indicator. NO Districts District Area (km2) NO Districts District Area (km2) 1 Saakow 6774 9 Bulo Burto 16054 2 Jowhar 4637 10 Baardheere 15275 3 Balcad 4238 11 Jalalaqsi 3123 4 Jalalqsi 3123 12 Kurtunwaarey 2546 5 Luuq 8260 13 Qansax Dheere 3267 6 Belet Weyne 14867 14 Rab Dhuure 3309 7 Bulo Burto 16054 15 Tayeeglow 6596 8 Garbahaarey 8318 16 Doolow 1618 https://doi.org/10.1371/journal.pone.0304202.t001 Fig 4. Sentinel-1 imagery acquired over Belet Weyne city in Somalia on 30th May 2016. https://doi.org/10.1371/journal.pone.0304202.g004 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 6 / 27

large swaths of high geometrical and spectral performance with imagery available at the spatial resolutions of 10 m, 20 m and 60 m. Sentinel-2 is broken down into 100 km x 100 km side tiles, known as “granules”. In total, we selected 15 Sentinel-2 granules covering the 16 AOI dis- tricts. Fig 5 shows the distribution of Sentinel-2 granules over the study area. On average, we downloaded 240 Sentinel-2 images for each granule from Google Cloud at the Level-1C processing level for a period of January 1st, 2016, to December 31st, 201, covering an area of 180,757 km2. In total, we downloaded 4169 of Sentinel-2 granules that occupied 2.5 TB of memory. We automatically pre-processed these datasets using Sen2cor (version 2.8) [27] to convert to Level-2A ready for processing. The Sentinel-2 datasets were used in several envi- ronmental indicators such as spectral vegetation indices, drought, flood, and Land Cover (LC) change. In addition, Sentinel-2 was used in flood mapping to visually verify the location and date of the flood events with an independent dataset. NASADEM. The NASADEM is currently the Digital Elevation Model (DEM) with the highest resolution that is freely available for Somalia. NASADEM was derived from the re-pro- cessing of the Shuttle Radar Topography Mission, using the additional layers from Terra Advanced Spaceborne Thermal and Reflection Radiometer, Global Digital Elevation Model, Version 3 data, Ice, Cloud, and Land Elevation Satellite, Geoscience Laser Altimeter System

ground control points of its LiDAR shots to improve surface elevation measurements that led to improved geolocation accuracy. We automatically downloaded NASADEM from the Land Processes Distributed Active Archive Center [28] in “.hgt” format distributed in 1˚ x 1˚ tiles (available for all land between 60˚ N and 56˚ S latitude), WGS84 for the year 2000 at approxi- mately 30 m pixel spacing. A total of 45 granules were used to generate a DEM mosaic cover- ing all the Sentinel-1 tiles. Other re-processing improvements include the conversion to geoid reference and the use of GDEMs and Advanced Land Observing Satellite Panchromatic Remote Sensing Instrument for Stereo Mapping AW3D30 DEM, and interpolation for void filling. The NASADEM was used to generate the following ancillary layers: a DEM mask, slope mask, and Height Above Nearest Drainage (HAND) index mask [24]. Soil moisture index (SMI). Soil moisture data represents the current state-of-the-art sat- ellite-based soil moisture data record production, in line with the “Systematic observation requirements for satellite-based products for climate” defined by Global Climate Observing Fig 5. The footprint of the Sentinel-2 granules acquired from January 2015 to December 209 covering the entire of the study area. https://doi.org/10.1371/journal.pone.0304202.g005 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 7 / 27

System. The SMI offers a spatial resolution equal to 0.25-degrees and with a 10-day temporal resolution. We acquired SMI for the study area for the period between 1978 and 2019 at the ESA Climate Change Initiative Soil Moisture Climate Copernicus. The soil moisture data pro- cessed into SMI anomaly used to develop the agricultural drought indicator. Standardised precipitation-evapotranspiration index (SPEI). SPEI is a metric that delineates the climatic water balance, synthesising the variance between precipitation and potential evapotranspiration over diverse temporal scales [29]. Offering near real-time insights into hydric deficit or surplus, SPEI enables monitoring with a monthly cadence and 1-degree spatial resolution. While SPEI-1 delineates short-term, monthly meteorological anomalies, sensitive to abrupt climatic shifts such as sporadic rainfall events, Conversely, SPEI-3 encom- passes a tri-monthly scope, yielding a more stable reflection of meteorological trends by miti- gating the influence of transient climatic fluctuations. SPEI-1 and SPEI-3 data spanning from 1950 to 2019, sourced from the Spanish National Research Council [26], indicate SPEI’s dynamic range from -5 (intense dryness) to 5 (significant wetness). Specifically, an SPEI value below -1 signifies a moderate drought, whilst a reading below -2 is indicative of an extremely drought scenario. IDP data. We compiled data from UNHCR, IDMC, and the World Bank where district- level statistics are available. The specifics of these derived and collected indicators are detailed

in the ensuing section. As part of the non-environmental indicators, we used IDP data (sourced from UNHCR) to study the internal migration of environmental refugees since 2015. These IDP figures represent the number of internally displaced people each month and include the cause of migration and the originating and destination districts. Socio-economic indicators. Data availability and quality present significant challenges for socio-economic indicators in Africa, particularly in impoverished countries like Somalia [30–32]. This issue is further exacerbated when seeking disaggregated data for smaller admin- istrative units. While international databases developed by various institutions cover diverse domains such as democracy, development, and access to services [33], they mostly provide annual statistics at the country or regional level, with limited availability of data for smaller administrative units. To address this challenge, our first step involved selecting data sources that provided disag- gregated socio-economic indicators at the district level on a monthly basis. We evaluated data from the World Bank, the International Organization for Migration (IOM), and the UNHCR to assess their quality and completeness for the study period from 2016 to 2019. Among these sources, we found that the most sufficient data is related to food prices at the district level on a monthly basis, providing information on the prices of essential commodities such as sorghum, maize, and rice. It is essential to note that while we acknowledge the limitations, inconsistencies, incom-

pleteness, and scarcity of data, the socio-economic variables were not selected as indicators for migration drivers. Instead, they were utilised to control for the impact of environmental fac- tors on the socio-economic conditions in Somalia. By including these variables, we aimed to contextualise the influence of environmental factors on the broader socio-economic landscape of the region. Methodological approach: Developing environmental indicators A series of environmental and non-environmental indicators were considered to understand environmental migration better. The goal of using environmental indicators was to measure the direct impact of environmental events on agriculture, human settlements, and public infra- structures, whilst non-environmental indicators such as IDP were used to quantify the human PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 8 / 27

impact. The environmental indicators were viewed as casual indicators, while non-environ- mental features measured the effect these causes had on the local population that ultimately leading to migration. Methodological approaches for each environmental indicator were developed and explained in detail. Agricultural drought indicator (ADI) A methodology was developed for ADI, drawing on the cause-effect relationship for agricul- tural drought. This relationship posits that a shortage of precipitation leads to a soil moisture deficit, which in turn results in a reduction of vegetation productivity (see Fig 6). The method- ology combines anomalies of precipitation (SPEI-1 and SPEI-3), soil moisture (e.g., SMI) and vegetation stress (e.g., NDVI). Here is how NDVI calculated in this study: NDVI ¼ NIRðband 8Þ Redðband 4Þ NIRðband 8Þ þ Redðband 4Þ ð1Þ • SPEI-1 and SPEI-3: to identify the precipitation deficit. • SMI anomaly: to identify soil moisture anomalies. • NDVI anomaly: characterise the subsequent effect of soil water stress on crops. As illustrated in Table 2, a rainfall deficit (watch level) causes a reduction of the soil mois- ture (warning level), and a water deficit in the ground, leading to vegetation stress. This stress results in a reduction of NDVI values (alert level) which ultimately impacts agricultural pro- duction and could lead to food insecurity. The ADI developed for this study is a type of Com- bined Drought Indicator (CDI) derived by combining the SPEI, Soil Moisture Index anomaly

(SMIa), and NDVI anomaly (NDVIa). SMIa is derived from anomalies of the SMI archive data Fig 6. Mosaic of Agricultural Drought (ADI) map of the study area in Somalia with 20m resolution in September 2017. The extent coordinates for AOI: 41, 0.3 to 47, 5.4. https://doi.org/10.1371/journal.pone.0304202.g006 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 9 / 27

between 1992 and 2018. SMIa ¼ SMI SMIðavgÞ SMIðstdÞ ð2Þ SMI(avg) represents the long-term average soil moisture and SMI(std) corresponds to its standard deviation and SMI represents one soil moisture image corresponding to a specific date. NDVIa is an indicator of access vegetation health and productivity, and it represents deviations of the NDVI from the long-term mean values. NDVIa ¼ NDVI NDVIðavgÞ NDVIðstdÞ ð3Þ The expressions NDVI(avg) and NDVI(std) denote the long-term averages and standard deviation, respectively. Given the non-existence of the Sentinel-2 data prior to late 2015, the values for NDVI(avg) and NDVI(std) were derived from the Sentinel-2 archive spanning late 2015 to December 2019. It is imperative to note that the utilisation of Landsat data to augment the archive was not deemed viable. This decision was underpinned by the discrepancies in spa- tial and spectral band configurations between the Sentinel-2 and Landsat datasets, making them incompatible for cohesive analysis. Flood. The established methodology for flood mapping capitalises on the variation in backscatter intensity discerned from pre- and post-flood SENTINEL-1 SAR imagery. The SENTINEL-1 imagery was pre-processed using SeNtinel’s Application Platform (SNAP) Tool- box 7.0 to carry out the following steps: (a) Apply orbit file, (b) Ground Range Detected (GRD) Border Noise Removal (border limit: 5000, trim threshold: 1.0), (c) Thermal Noise removal, (d) multi-looking (2 looks in range and 2 in azimuth), (e) speckle filtering (Refined

Lee filter which is an enhancement of the Lee filter in order to provide better preservation of edges, linear features, and texture information, 5x5 window), (f) range Doppler terrain correc- tion using SRTM 1arc sec and geo-coding to WGS84/UTM Zone 38N. The VV polarisation was considered to detect flooding as opposed to HH or VH polarisation [34, 35]. Intriguingly, full-polarisation SAR data extends the capability to apply advanced polarimetric techniques, such as ’target decomposition’, which could potentially refine object detection further than raw polarised data alone. While beyond this paper’s scope, future research could harness these techniques to heighten accuracy, especially in discerning specific objects within flood contexts. Subsequently, pixel-based change detection between the pre-flood event at time (t1) and the post-flood event at time (t2) was performed using the intensity ratio: R ¼ It1ðiÞ It2ðiÞ ð4Þ Where "It1(i)" is the backscatter intensity of pixel "i" in image acquired at “t1”, and "It2(i)" is the corresponding intensity pixel in image acquired at time "t2". The pixel ratio R(i) is a sample of a random process whose Probability Density Function (PDF) for uncorrelated SAR Table 2. ADI classification scheme with three drought levels. Drought level Conditions Interpretation Watch (yellow) SPEI-3 < -1 or SPEI-1 < -2 Low SPEI means precipitation deficit. Warning (orange) SMI < -1 & (SPEI-3 < -1 or SPEI-1 < -2) Precipitation deficit plus a negative anomaly of soil moisture. Alert (red) NDVIa < -1 & (SPEI-3 < -1 or SPEI-1 < -2)

Precipitation deficit plus a negative anomaly of vegetation ‘greenness’. https://doi.org/10.1371/journal.pone.0304202.t002 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 10 / 27

intensity images is [35]: PðRÞ ¼ Gð2LÞgLRL 1 G2ðLÞðg þ RÞ2L ð5Þ Where γ is the true change of the radar cross sections σ0, and L is the number of looks. The probability of detection Pde given a threshold T, L and γ is: PdeðR > TÞ ¼ Z 1 PðRÞdR ð6Þ It is important to notice that P depends on the number of looks L. To illustrate the impact of L on the detector performance let us compute the Pde for the detection of a 3 dB change (γ = 0.5) with T = 0.4. For L = 4 (Sentinel-1) we have: Pde = 0.58, for L = 50 (Lee filter) Pde = 0.86. These figures point out the importance of speckle filtering for the ratio detector. In our imple- mentation the threshold T is determined by supervised learning, i.e., by estimating the ratio in a known area of no change, as it will be described in Thresholding section. This approach minimises the misclassification of arid land as flooded due to similar back- scatter of desert surfaces and flooded areas in arid regions like Somalia. Apart from common thresholding approaches were tested (e.g., unimodal thresholding approaches such as T-point and maximum deviation), a change detection approach was employed based on a ratio between pre-flood and post-flood imagery. This change detection strategy is designed to mini- mise the misclassification of arid landscapes as inundated terrain a frequent challenge due to the comparable backscatter characteristics of desert and water surfaces in arid zones such as Somalia. The current method’s reliance on VV polarisation is rooted in established efficacy,

yet it is recognised that incorporating target decomposition features, which utilise the compre- hensive data from all polarisations, could substantially mitigate this issue. Such an advanced technique promises a refined analysis, leveraging the full spectrum of polarimetric data to dis- tinguish between dry land and aquatic environments with greater precision. The thresholds in this study were calculated by supervised inspection of the imagery. The optimal threshold was calculated for a flooded pixel (PF) and for a non-flooded pixel (PNF). A coefficient fc was calcu- lated based on several trial-and-error tests and also confirmed by findings in [36]. The μ and σ are the mean and standard deviation of the no-change area selected in the ratio image R: PF � m½R� þ fc∗s½R� ð Þ PNF < m½R� þ fcs½R� ð Þ ð7Þ Where PF is the flooded pixel; PNF is the non-flooded pixel; fc is a coefficient factor found based on several iterations; μ = mean R and σ = the R standard deviation over the stable area. Next, ancillary masks were generated based on NASADEM. A threshold was set to generate a binary slope mask. Slopes � 3˚ were considered to be prone to flooding and assigned a pixel value of 1, while slopes >3˚ are considered not likely to flood, assigned a pixel value of 0 and were removed from further analysis. This step was crucial in eliminating any pixels that may have exhibited a change in backscatter due to the angle of signal return from a higher elevation such as hills. Areas with an elevation � 200 m were also masked and deemed unlikely to be

flooded. The slope and elevation threshold were selected based on the values used in the litera- ture (e.g., [34, 37]). Finally, a further terrain filter was applied to remove areas that can be considered unlikely to flood based on the Height Above Nearest Drainage (HAND) [38]. A binary HAND mask was generated where pixels with HAND values �15 m were considered not flood-prone while pixels with HAND < 15 m were considered likely to be flooded. Post-processing of the result- ing Sentinel-1 flood map was performed to remove isolated clusters of pixels < 50 pixels in PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 11 / 27

size (equivalent to 0.25 ha), reducing the false detections arising from noise. The algorithm removes and replaces pixels below the selected threshold with the pixel value of the largest neighbour polygon. Finally, a morphological binary closing was applied to fill holes smaller than the chosen disk radius (r = 5) [39]. Land Cover (LC) change. A LC classification scheme was developed to study the impact of the natural disasters like drought and flood on agricultural food productivity and human activity. This motivated the inclusion of the cropland class, containing food producing arable land as well as food producing orchards. Meanwhile the Settlement class was created to include any apparent permanent human structure. The two principles “background” classes in this product where the Grassland/Shrubland class and the Soil/Sparse vegetation class. The pri- mary differentiator between these two classes is the quantity of vegetation present, with Grass- land/Shrubland dominated by low bushes, and occasional thick grass coverage, whilst the Soil class contains minimal vegetation. The study area is covered mostly with grasslands and shrub- land whilst the majority of the croplands, forest, and settlement regions are distributed along the Shebelle River. The classification scheme can be seen below in Table 3: The entire of four spectral bands of Sentinel-2 with 10m resolution (blue, green, red, near infrared) and two 20m resolution (Short Wave Infrared 1 and 2) were used at the Level-2A in

the training and the evaluation of the model. Additionally, the NDVI was used to enhance the distinction between the cropland and forest classes. A U-Net based model was chosen due to its superior performance in preliminary studies, both in terms of validation metrics and visual results. U-Net architecture is a type of Convolu- tional Neural Network (CNN) with layers that “slide” multiple trained filters to the image, cre- ating several output channels. In U-Net, all convolutions have a 3x3 kernel size, meaning that a pixel and its immediate neighbours are all selected. The CNN implemented using Python 3 programming language. The encoding part of the network used skip connections to form so called ResBlocks, which advantageously allow the input image to propagate deeper into this section of the network, thereby preventing issues such as vanishing gradients from being prominent. The loss function used in the network training was a weighted average of the cross-entropy and dice loss between the two images, with the dice cross comprising 90% of the loss function and the cross-entropy the remaining 10%. The network was trained using the Adam Optimiser combined with a learning rate sched- uler, which gradually reduces the learning rate throughout training by a constant factor of 0.92 per epoch. The network was trained for 100 epochs in total. The U-Net model had various hyperparameters which affected the details of the network architecture and training Table 3. Classification scheme of AOI land cover with six members. ID Class Description 0 Agricultural Fields (green)

Arable land used for crop growing. Food producing orchards are also including in this class. 1 Forest (dark green) Area covered by dense tree coverage, such that the underlying ground coverage cannot be seen. 2 Grassland / Shrubland (brown) Open areas covered by dense non-arable vegetation including predominately from grass and bushes. Class may also include sparse tree coverage. 3 Settlement (grey) Region forming a permanent human settlement. 4 Soil / Sparse vegetation (yellow) Open area covered by bare soil, sand and low-density vegetation. 5 Water (blue) Water bodies including lakes, rivers, ponds, and canals. Also includes regions of flooding covered by ground water. https://doi.org/10.1371/journal.pone.0304202.t003 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 12 / 27

parameters that needed to be tuned in order to get an optimal result. There were two main hyperparameters that defined the structure of the U-Net, its depth, and its starting features. The depth of the U-Net is defined as the number of skip connections between convolutional encoding/decoding blocks. The depth of the model is 5, as there are 5 total skip connections between encoder/decoder blocks. The other key hyperparameter in the structure of the net- work was the number of starting channels that were used. A depth of 5 was shown to give the highest accuracy and F1 score on the validation set and was able to fit in the GPU memory. For the number of starting channels, as value of 64 was chosen by the same process as before of experimenting with values from 16 to 128, with 64 giving the highest accuracy on the validation. For training and testing the U-Net model, a series of polygons were created using Sentinel- 2 imagery. The polygons, selected during the two annual growing seasons, were devoid of any distortions. The U-Net model was trained on 128x128 images containing the seven input channels. The first approach of gathering training samples, inspired by an approach [40], syn- thetic training images were created from the training polygons. The algorithm to generate the synthetic training images via tessellation worked by randomly selecting a polygon from the training polygon set and then rasterizing it which involved calculating all input features. Each rasterised polygon was sampled randomly weighted by its area, meaning that larger polygons

were more likely to be selected than smaller ones. In total, 1024 128x128 training images and 128 testing images of 128x128 were created. The second approach to generating training images for the U-Net was to select a square ground region and attempt to label each pixel in the image using the Sentinel-2 data as a refer- ence. Each pixel in an image was hand labelled into one of the six classes defined in the classifi- cation schema. After augmentation this led to a total of 985 ground truth images of 128x128 in size, generation in total 2009 training images from both methods available for training. These polygons were then divided with 70% of the polygons being used for training and 30% for testing. The LC Change was calculated based on the detection of changes between two LC maps: a before and after map. The LC change between the maps can then be calculated by performing a simple mapping of the form: LCChange ¼ LCafter þ LCbefore ∗ #classes ð8Þ The number of classes was 6 resulting to a total of 36 possible LC Change classes represent- ing each possible transition of any class to any other class (including conservation). In the case where one or both LC values were unavailable for a given pixel due to the cloud mask, then the LC Change at this pixel was also masked. All possible transitions are seen in Table 4. Analysis and results Drought and floods In total, 48 ADI mosaic were produced monthly with 20 m spatial resolution in GeoTIFF for- mat. Fig 7 shows an example of drought map from August 2017, where the AOI was predomi-

nantly affected by moderate and severe drought, especially in the south near the lower Shebelle River. River floods, caused by water overflowing riverbanks due to prolonged rainfall, are a main cause of environmental migration in Somalia. A flood algorithm was developed using a change detection technique with Sentinel-1 data from before and after a flood. The Digital Elevation Model was used to calculate the slope of the study area. Unlike the monthly drought maps, flood maps were produced based on flood events. Fig 8 displays the overall flood mosaic map. PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 13 / 27

In 2016, 865.5 km2 of the AOI flooded, primarily affecting the Belet Weyne, Saakow and Jowhar districts (Fig 7A). In 2017, floods affected the districts of Saakow and Jowhar, covering 2263.4 km2 (Fig 7B), whilst in 2018 several districts were flooded (including Hiiraan -Fig 6) with an area equal to 3552.1 km2 (Fig 7C). Lastly, in 2019 the total area flooded was 2657.9 km2. The flood was used as a base map alongside the population map to estimate the number of people affected by each flood event. Land Cover (LC) and LC change maps for drought and flood events The final environmental indicator comprised of the LC and LC change maps, which are crucial for assessing the impact of natural phenomena like floods and drought on agricultural food productions (like orchards and croplands) and human settlements (like towns and villages). First, LC maps were produced to show the Earth’s different land cover patterns in five different classes (soil/sparse vegetation, water, agriculture field, forest, and grassland/shrubland). Sentinel-2 Level 2A imagery was the primary data input for LC maps. A cutting-edge U-Net based algorithm produced LC maps during drought and flood events, enhancing detec- tion of different textural features seen in the base images. LC is a fundamental environmental variable for understanding the causes and trends of human and natural processes. LC refers to the physical material at the earth’s surface like grass, soil, impervious surfaces, and trees. The information extracted from the land cover features and their associated changes can help

Table 4. Demonstration of the mapping of the transition between two land cover classes to an integer value that was stored in the land cover change raster file. To From Agricultural Field Forest Grass/ Shrub Settlement Soil Water Agricultural Field 0 1 2 3 4 5 Forest 6 7 8 9 10 11 Grass/Shrub 12 13 14 15 16 17 Settlement 18 19 20 21 22 23 Soil 24 25 26 27 28 29 Water 30 31 32 33 34 35 https://doi.org/10.1371/journal.pone.0304202.t004 Fig 7. Annual aggregated flood maps of the study area generated from Sentinel-1 data. a. 2016 floods; b. 2017 floods; c.2018 floods; d.2019 floods. https://doi.org/10.1371/journal.pone.0304202.g007 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 14 / 27

understand the impact of environmental threats on the amount of variation of crops. Accurate identification and extraction of different land cover features (e.g., agricultural fields) is impor- tant for assessing distribution across agricultural regions. The Table 5 describe the statistical metric for the accuracy of the land cover model. While the accuracy, precision, and recall on the test set provide an insight into the product’s potential accuracy, a more rigorous accuracy assessment was carried out to obtain a closer esti- mate of the genuine accuracy. Random stratified points were chosen across the entire AOI, and very high-resolution imagery (e.g., Google Earth) was employed to determine the actual ground truth class for each point. The overall accuracy of the aggregated LC maps was esti- mated at 78%, a decline from the generalisation accuracy of 92% derived from the unseen test set. This decreased performance is likely attributable to the substantial amount of new, unseen data introduced to the model. Moreover, this dip in overall accuracy could be mitigated by referencing VHR imagery from Google Earth as ground truth, especially in instances where the Sentinel 2 imagery was indistinct or inconclusive. In response, 17 LC maps were produced corresponding to significant flood events leading to IDPs from 2016 to 2019. The dates for creating LC maps were selected using the first avail- able Sentinel-2 image after for each flood event. LC maps for drought events (see Fig 8) were generated as aggregates for growing seasons; Gu from March to late May and Deyr from the

end of September to the end of December). Eight aggregate LC maps were produced, provid- ing a general representation of Land Cover for each agricultural growing season between 2016 Fig 8. Aggregated land cover map of the study area represent the Deyr growing season generated from Sentinel-2 time-series from 1st October to the end of the December 2018. https://doi.org/10.1371/journal.pone.0304202.g008 Table 5. Precision, recall and F1 scores as calculated on a prediction of a mosaic comprised of single Sentinel-2 images from November 2017. Class Precision Recall F1 Score Support Agricultural Field 59.4% 62.3% 60.8% 61 Forest 47.8% 33.7% 39.5% 98 Shrubland/Grassland 60.3% 80.7% 69.0% 243 Settlement 72.7% 88.9% 80.0% 18 Soil/Sparse Vegetation 79.3% 48.3% 60.1% 151 Water 94.5% 92.9% 93.7% 56 https://doi.org/10.1371/journal.pone.0304202.t005 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 15 / 27

and 2019. The Fig 8 shows the aggregate LC map on Deyr 2018. The Agricultural fields shows as light green colour represents the croplands and orchards mainly distributed across the She- belle and Juba rivers. Following this, LC change maps were produced to highlight changes in agricultural fields and human settlements due to flood and drought events. LC change maps for flood events were generated by comparing post-flood LC maps to corresponding aggregate LC maps. Simi- larly, for drought, LC Change maps were produced by comparing each aggregate to the previ- ous years’ aggregate map for the same season; for example, Gu 2016 was compared to Gu 2017. Fig 9A presents the aggregated Land Cover (LC) map for Deyr 2019 over the city of Beled- weyne. A series of LC Changes were determined by comparing each LC map of the Gu and Deyr with the subsequent year’s LC maps. These LC Change maps highlight the effects of drought and flood on settlements and agricultural fields. All the LC and LC change maps were produced with a 10m spatial resolution. The flood LC Change map in Fig 9B shows the clear impact of floods on the agricultural fields and settlements of the affected areas, whilst Fig 9C and 9D demonstrate the impact of the flood by showing the agricultural fields and settlements that were lost in the Beledweyne District, respectively. All the LC and LC change maps were produced in 10m spatial resolution. . . Most notably, evidence of migration from major urban areas close to the rivers to smaller communities far away from rivers can be seen after flood

events. This suggests that migrants from major floods try to stay close to their homes but out- side the direct disaster zone. An unsupervised clustering was performed to further explore the impact of flood events. Regular flood events exhibits similar characteristics, whereas extreme floods led to higher IDPs, more affected people, and caused larger damage to agricultural fields and human settle- ments. It is expected that, compared to regular floods that are part of the agricultural climate and essential for the growing season, extreme floods are more likely destructive and lead to migration. Compared to flood events, drought events typically take place over a much longer time scale. As such for determining the LC change, a single post drought image cannot be defined in the same way as for flood. As a result, six drought LC change maps were produced based on Fig 9. Flood maps of the study area. a. Aggregated LC map in Deyr 2019; b. LC Change map on flooded areas of Beledweyne after flood event (No 11); c. LC Change map with drought impact on agricultural fields; d. LC Change map with the impact of flood event (No 11) on settlements area. https://doi.org/10.1371/journal.pone.0304202.g009 PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 16 / 27