Satellite Analysis of the Environmental Impacts of Armed-Conict in Rakhine, Myanmar

The impacts of armed conict on the environment are extremely complex and dicult to investigate, given the impossibility of accessing the affected area and reliable data limitation. Very-high-resolution satellite imageries and highly reliable machine learning algorithms become very useful in studying direct and indirect impacts of war on the ecosystem, in addition to connected effects on human lives. The Rohingya conict is described as one of the worst humanitarian crises and human-made disasters of the 21st Century. Quantication of damage due to the conict and the suitability of human resettlement has been lacking despite the ongoing agreements to repatriate refugees and the importance of ecosystem services for the communities' survival. Here we report the investigation of environmental conditions pre-, during, and post-conict in the conict zone using satellite data. We implemented and experienced the Google Earth Engine (GEE) cloud-based computing platform with a widely applied algorithm, the Random Forest (RF) classier. We found striking near-complete demolition of inhabited regions, dramatic and highly signicant increase in burning areas, and substantial deforestation. We discuss the reasons behind such ndings from the Rakhine case and debate some general conservation lessons applicable to other countries undergoing post-conict transitions. eld geoeld high-resolution our eld conict to ongoing conicts and travel restrictions. In of ground-truth data, I derived an extensive training dataset from expert interpretation from a set of VHRI Pléiades satellite data. our classication images are VHRI, this method offers better accuracy than using Google Earth Pro for collecting training and validation data. Based on experts' regional knowledge of LULCC in the study area, I manually delineate reference data through visual interpretation, producing datasets for each study period. increased environmental pressure in Myanmar have been documented as large-scale development projects and shifting cultivation, the underlying drivers are deeply embedded within the country’s socio-economic and political context. Land alteration during and after civil war are also recognized as the crucial drivers of environmental disasters, albeit they are still poorly understood phenomena in Myanmar. A critical analysis of armed conict's environmental impacts provides vital information on ecosystem damage triggered by conicts and the mitigation strategies. Our study examines the spatial and temporal pattern of environmental changes and the complex dynamics of land-cover-land-use shifts in war-torn Rakhine. By commissioning very-high-resolution-satellite-imageries and machine learning techniques, I set out to evaluate the capacity of very-high-resolution-satellite data to quantify the magnitude of damage in conict zones where the study context is notoriously dicult to assess. Remote sensing allowed for a comprehensive analysis of forest and land cover conversion in the Rohingya conict zone in a 28590.724 ha study area in the Rakhine State of Myanmar from 2012 to 2019. Analysis reveals that the environment and ecosystem have undergone unprecedented change and have witnessed a dramatic forest cover decline. The region experienced a horrendous loss of residential area (nearly 100% wiped-out) regarding human life and livelihood loss. LULCC included an appalling expansion of burned areas and a barren and scrubland, resulting in a drastic reduction in all other forms of land cover, forest, wetland, human settlement, development, and agricultural lands. Forests predominantly covered the land, wetland ecosystems, planted/cultivated land before the conict. However, all these land cover classes dropped off to virtually nil during the conict, indicating a linear relationship between the environmental and human crisis and the conict.

Time-series of satellite imagery are useful for monitoring the ground's situation over a long period to record changes resulting from con icts (Machlis and Hanson 2011;Weir, McQuillan, and Francis 2019). I use satellite images from 2012, 2017 and 2019, to quantify the environmental degradation and livelihood destruction between pre-con ict and during the con ict in Maungdaw and Buthidaung -situated in the northern tip of Rakhine state, the epicenter of the con ict. Some of the most visible impacts of armed con icts are the imprints left on the landscape. Remotely sensed very-highresolution (VHR) satellite data could be very useful in detecting environmental changes and depicting destroyed communities' livelihood in extremely sensitive con ict-zone. This research is motivated by the fact that the environment continues to be the silent victim of armed con icts, and con ictinduced environmental degradation causes signi cant harm to the communities that depend on their livelihoods on the ecosystem services. The paper addressed six critical questions: (1) What is the intensity and pattern of environmental degradation in this con ict zone? (2) What is the magnitude of human life and livelihood destruction? (3) How many forests have been lost during the study period? (4) Are there any post-con ict afforestation, and what is the pattern of afforestation in the study area? (5) Is the place suitable for human resettlement at this current state? Finally, (6) What is the reliability of satellite data and machine learning in assessing complex con ict zones.

Study Region
The study area is located in the northern part of Myanmar's Rakhine State in Myanmar's westernmost part. It is located approximately between latitudes 17°30' north and 21°30' north and longitudes 92°10' east and 94°50' east. The total area of the state is 36,769 Km 2 . Rakhine is one of the deforestation hotspots and has the second-highest forest loss in Myanmar. Rakhine region lost 109,024 ha of intact forests in between 2002-2014.
Rakhine also lost a signi cant area of mangrove forests with the highest annual loss rate in Myanmar (De Alban et al. 2020;Richards and Friess 2016). Consequently, the Rakhine state became the country's most vulnerable state to natural disasters (Kissinger et al. 2017). The contributing factors to land cover change or deforestation in Rakhine state are currently unknown and not well represented in the current literature, albeit having alarming levels of con icts and displacements (Lim et al. 2017).
The Rohingya con ict is characterized by sectarian violence between the Rohingyas and ethnic Rakhine communities and a military crackdown on Rohingyas by the security forces (International Crisis Group 2019). There are also ongoing clashes between the Myanmar military and the ethnic Rakhine Arakan Army (HRW 2020). The satellite images of the study site cover 28,590.734 ha of con ict-affected areas in Maungdaw and Buthidaung townships. The study area was chosen based on the severity of con ict impacts and the high-resolution satellite data availability. The con ictaffected areas were identi ed based on Human Rights Watch's damage analysis reports (HRW 2017). HRW released the location maps of con ict extent and damage zones in Maungdaw and Buthidaung regions using high-resolution satellite images. Our satellite data covers most of the damage zones identi ed by HRW. The total study site's extent is 2,326.5615 km at the top, 2,299.757 km at the bottom, 415.431 km to the left, and 440.001 km to the right. Geographically, the area examined in this study extends from 19° 49'41.63" N (elevation 64m) to 19° 52'26.77" N (elevation 198m) latitude and 94° 03'20.57" E (elevation 333m) to 94° 06'42.03" E (elevation 133m) longitude. The region supports tropical rain forests, tropical mixed evergreen and deciduous forests, coastal mangrove swamps, and inland swamp forests (Davis 1965). The region is disaster-prone and has very low food security due to low agricultural productivity.

Data And Image Processing
Mapping with complex machine learning classi ers requires many representative datasets to train and validate the models (Nomura and Mitchard 2018). Figure 2 below depicts the work ow for the classi cation of satellite images and producing the maps. The representative training samples can be usually collected from the eld to obtain geo eld photos or high-resolution images (Nomura et al. 2019). In our case, conducting eld visits to con ict zones was impossible due to ongoing con icts and travel restrictions. In the absence of ground-truth data, I derived an extensive training dataset from expert interpretation from a set of VHRI Pléiades satellite data. As our classi cation images are VHRI, this method offers better accuracy than using Google Earth Pro for collecting training and validation data. Based on experts' regional knowledge of LULCC in the study area, I manually delineate reference data through visual interpretation, producing datasets for each study period.
I generated a total of 2,000-3,682 polygons for each respective year distributed throughout the study area to cover the entire satellite image and carefully digitized for training and validation data set. Land cover classi cation of the study area was primarily determined based on the existing land cover analysis and classi cation maps of Rakhine state developed by the United Nations Operational Satellite Applications Programme (UNOSTAT) in 2015 (OCHA 2015). The maps provide land cover classi cation over Rakhine state derived from Landsat 8 multi-spectral imagery acquired between January to February 2015 at 30m pixel resolution. The classi cation is divided into ve main classes: Forest, Mangrove, Cropland (Paddy Field), Barren Soil, and vegetation. For this study, using VHRI data, I classi ed a total of eight categories: Forest, Barren, Development, Planted/Cultivated, Scrubland, Water, Wetland, and Burned Area. These training samples and reference data were later used as input variables for the Random Forest model's calibration. At the rst step of classi cation, I created a binary column in the training dataset to split polygons into train (50%) and test (50%) using R-studio programming software.
To quantify Land-Use-Land-Cover-Change (LULCC) associated with the armed-con ict, primary data from the existing time series from the commercial VHRI orthorecti ed satellite images, Pléiades -1A Satellite Sensor, were collected (AIRBUS 2020). Pléiades 1A was the rst high-resolution satellite in the Airbus Defense and Space constellation. Data captured very close dates for three years in the same study area. To compare pre-con ict, con ict and post-con ict periods, images from three acquisition dates (t0=2012 November 11, t-1= 2017 November 28, t-2= 2019 November 26) with a spatial resolution of 0.5m (multi-spectral bands) were used for our main study area in Maungdaw and Buthidaung townships. These images were mainly used to track and classify land cover changes in the con ict zones, using the RF classi cation method. Pléiades -1A images are multi-spectral with four bands (blue: 430 -550 nm, green: 500 -620 nm, red: 590 -710 nm, near-IR: 740 -940 nm). The (pre-, during, and post-con ict) dates were chosen based on con ict peak times and cloud-free images for the study areas. Due to the monsoon climate patterns throughout Myanmar, it is challenging to obtain cloud-free satellite imagery. There are also no Synthetic Aperture Radar (SAR) covering the study period in our study region. All the images were pre-processed and mosaicked in ArcGIS Pro.
Next, I used Google Earth Engine (GEE), a cloud-based computing platform, to perform a machine learning classi cation algorithm. GEE is a cloudbased platform for planetary-scale geospatial analysis that enables Google's massive computational capacities to unravel the most signi cant problems concerning the land cover mapping of large geographical areas and big data (Gorelick et al. 2017;Noi Phan, Kuch, and Lehnert 2020). Due to its cloud computing power, data processing works, and the memory capacity of the user's computer is not a limiting factor when working with big data and imagery. Another big advantage of GEE is the consolidated library of low to moderate resolution global satellite data to supplement key VHRI data (Sidhu, Pebesma, and Câmara 2018). GEE is an e cient platform to execute complex work ows of satellite data processing required for largescale applications such as LULCC monitoring (Shelestov et al. 2017). Figure 2 below summarizes the work ow of our study using GEE, R studio, and ArcGIS Pro. I rst uploaded the main satellite data and training data in Google Cloud storage and ingested into GEE asset using Python application programming interface (API). Subsequently, I extracted training and testing data in GEE to update training parameters and testing data size, naming the datasets and matching bands. I then added new bands for additional exploratory variables to include the digital elevation model (DEM), slope, and vegetation indices. I obtained Sentinel-2 images from the nearest date of the original images from the GEE archive and calculated Normalized Difference Vegetation Index (NDVI) using following formula: I rst employed two supervised machine learning models Random Forest (RF) and Support Vector Machine (SVM). As the Random Forest RF model performed better than SVM, RF was chosen as our classi cation model after the classi cation. The highest overall accuracy obtained with SVM model was 46%. RF is a robust model particularly appropriate for LULCC classi cation as it can effectively process a large number of predictor variables and complex datasets (Bricher et al. 2013;Cutler et al. 2007). The RF model outperformed other traditional parametric based image analyses because of its capacity to conform missing values and complex variables and obtain high overall classi cation accuracy (Belgiu and Dra 2016;Sesnie et al. 2010).
RF creates an ensemble of trees, each providing a "vote" to select the best classi cation approach. The majority of votes from the assemblages of the tree created in RF determine the class assignment of the pixel, and the results of the large quality of trees are aggregated internally (Berhane et al. 2018). Before classi cation, I conducted RF hyperparameter tuning to update bands and select the best hyperparameter values. As it is necessary to have the speci cation of several parameters to execute the RF model, each RF tree was established by training each tree in the forest (ntree) with the number of input predictor-variables (mtry), randomly chosen at each split from the training dataset (Aung, Fischer, and Buchanan 2020). In the last section of RF classi cation in GEE, I set hyperparameters values after hyperparameter tuning, trained the classi er, and exported classi ed raster images and accuracy assessment results to the google drive folder.
Comprehensive geospatial information, such as the con ict zones' geographical location, road network, other physical features, affected villages, and the geographic boundaries of villages and townships were derived from the Union Enterprise for Humanitarian Assistance, Resettlement and Development in Rakhine (UEHRD), the databanks of the Humanitarian Data Exchange and Myanmar Information Management Unit's GIS resources.
In addition to the accuracy assessment estimated from RF classi cation in GEE during the bootstrapping process (Belgiu and Dra 2016), based on the best practices recommended by (Olofsson et al. 2014), I adopted a strati ed random sampling design. Accuracy assessment calculates the accuracy of maps, quanti es each class area, and evaluates the uncertainty of classi cations of the area (Mellor et al. 2013;Sharma, Hara, and Hirayama 2017). The required sample size was calculated using the following formula: I de ned a target standard error for the overall accuracy of 0.01. Using the proportional approach, I allocated a sample size of 50-100 to the smaller classes, and the rest of the samples were proportionately allocated for each change strata based on the class. The estimated variances are then computed based on the sample size allocation.

Results
The satellite imagery analysis results for the Rakhine con ict are striking (Table 1; Fig. 3 and 4). The classi cation maps of the study area from three points in time were investigated. This way, the environmental and human destruction in the con ict-affected zones were examined. The resulting maps are shown in Fig 4. Table 1 summarizes the calculations of the net change in the area of each land category.
The results show near-complete demolition of inhabited regions in the entire study area. At the original pre-con ict period, forests are the most extensive land cover area, followed by a residential area, water, and wetland, and planted or cultivated area. In 2012, forest covered over 27% of total land cover. The residential area covered 21% of the study region. Overall, in Maungdaw and Buthidaung (part of the ethnic cleansing), all the land cover categories experienced dramatic and highly signi cant alterations in the during-con ict years (Table 1; Fig. 3). First, the results reveal shocking evidence of the massive annihilation of residential areas and human settlements. The classi cation shows the decrease of 5897 hectares of human settlements, which is a horrifying 99.73% damage within three years, resulting in utter destruction. This is also manifested in the before and after maps (Fig.4), where burned areas (red) drastically expanded, and the area of human settlement (pink) become almost undetectable. This widespread devastation is the result of the burning of 6940 hectares in 2017 (Table 1). These results correspond to (Human Rights Watch 2017)'s report, which has documented arson, killing, and looting during the con ict. There is a small area of burned area in 2012. This can be partly due to the small-scale burning of villages during the rst wave of violence, which started in June 2012, and in some part, due to slash-and-burn during the winter seasons.
As a result of burning, there is also a substantial decrease in planted and cultivated lands. A similar pattern was observed in the forested area, showing an overall deforestation rate of 46.59%. There is also a notable decrease in the water/wetland area. This can be associated with the destruction of shponds, shrimp farms, and other aquaculture wetland ecosystems. Another area of signi cant depletion is development. Based on the study area's development characteristics, I speci ed only roads and large infrastructures as development areas. The result can be due to the deliberate destruction of road networks to prevent victims from eeing or returning. Hence, in the 2017 classi cation map, the development area (blue) is unapparent. Another appealing result is a sharp increase in barren and scrublands during the con ict. This result is also re ected in a drastic loss of planted/cultivated lands. In all, the results are elucidated in before and after maps where the vast majority of the land in 2017 is covered by burned area, barren, and scrubland (Fig. 3).
Our results from a post-con ict (2019) assessment of the study area reveal notable land cover changes (Table 1; Fig 3). There is a 97% increase in the residential area after losing more than 99% of the residential areas during the con ict. The net loss from pre-con ict periods remains over 90%. Highresolution satellite images revealed several new structures and settlements in the previously destroyed areas. There is a negligible decrease in barren/scrubland and an increase in water/wetland in the post-con ict period.
The study area's NDVI maps also support the quanti cation of environmental degradation in these trajectories. The NDVI maps for the three periods pre-, during-and post-con ict indicate that NDVI values signi cantly declined between pre-and during-con ict periods and continue to decline between during-and post-con ict periods but with a much lower rate of decline in the latter interval (Fig 5).

Discussion
The use of Very-high-resolution satellite data and machine learning algorithms to investigate the impacts of armed con ict revealed catastrophic and horrendous results. Over a span of eight years (2012-2019), the seven investigated land categories underwent substantial damage in the con ict zones. Primarily, the massive area of burned land resulted in sweeping destructions of human settlement, agricultural lands, forests, developments, and wetland ecosystems. From the pre-to during-con ict period, an enormous area of land became barren or scrubland. This is driven by extensive burning and agricultural land abandonment, often common in con ict settings and highly contested regions (Witmer 2008). The deforestation inside the con ict zone is immense and accelerated with the onset of con ict. Such a massive increase in natural habitat loss, often of primary forests, and livelihood destruction have profound effects on the ecosystem and human. These forests are important natural capitals and provide essential ecosystem services for the local communities. The report from (HRW 2017)'s pinpointed the near-total destruction of villages by re and reported that approximately 90-100% of all the villages are partially or destroyed. Our quanti cation and temporal comparison manifested more devastating and profound damage. The scale, scope, and timing of destruction also corroborate HRW's ndings. Consistent with reports of mortality, other crimes against humanity that forced Rohingya to ee for their lives, the current study found that a wide spectrum of heavy arson attacks was committed in the region. The results also revealed that the con ict's disastrous impacts are not just on the human lives and properties, also on the environment and ecosystem. In terms of environmental harm, the magnitude and intensity of damage are the most pronounced for agricultural lands and forests.
Although some recovery patterns were observed, the net decrease in both land categories remains signi cant in the post-con ict period. The ndings of previous works on environmental impacts and recovery during the post-war period have been contradictory (Gorsevski et al. 2012;Stevens et al. 2011;Suarez et al. 2018). Although the deleterious environmental impacts of war are felt long after the con ict in most cases, some studies suggested that some forests and biodiversity rebounded after the war due to a decrease in human pressure (Kaimowitz and Fauné 2003). In our case, the result obtained from the analysis opposes the narrative that war can have positive impacts on the environment, especially during post-con ict and peace periods (Clerici et al. 2020;Reardon 2018;Stevens et al. 2011;Zúñiga-upegui et al. 2019). Although the reduction in human pressure can be bene cial for the environment, it cannot guarantee sustainable environmental security. The formation of "no-go zones" due to compromised security (Gorsevski et al. 2012) is ultimately dysfunctional and symbolic of more deep-rooted problems. The disruption of state institutions, the collapse of non-governmental organizations active in environmental initiatives, poor management of natural resources, and discriminatory land law exacerbates them. Despite being replaced by cultivations and structures to some extent, the majority of the burned area also remain unchanged. The health and quality of vegetation also has not been improved until the end of the study period. Negligible change in barren and wetland is unsurprising as repatriation and resettlement of the victims have not been implemented until 2019, and economic activities such as aquaculture businesses in wetland areas are yet to revive.
These results con rm that armed-con icts and civil war aggravate Myanmar's already declining forest trend and degrading forest quality. Myanmar loses 0.87% of its forest cover annually (Aung et al. 2020;Leimgruber et al. 2005;Yang et al. 2019). Myanmar is also the current mangrove deforestation hotspot globally, and more than 90% of mangrove forests are located in Rakhine and Tanintharyi (De Alban et al. 2020;Richards and Friess 2016). Besides wood extraction, agricultural expansion and infrastructure development, civil war is recognized as a root cause of deforestation in Myanmar (Baskett 2016;Bhagwat et al. 2017;Connette et al. 2016;Lim et al. 2017;Prescott et al. 2017;Wang and Myint 2016). In this study, considering the intensity of forest losses occurring within the con ict zones' immediate proximity and during the con ict timeframe, all forest losses can be attributed to the con ict and its subsequent aftermath. While deforestation and afforestation co-exist, the area of deforestation is still more extensive than the afforestation. This result is different to ndings from other con ict zones, such as Nicaragua's Atlantic Coast, South Sudan-Ugandan border, and Congo (Gorsevski et al. 2012;Nackoney et al. 2014;Stevens et al. 2011). As planted/cultivated area is subsequently rebounding during the same period, the afforestation pattern might be contributed by agricultural expansion. Commercial agriculture has been used as opportunities by Military to con scate abandoned lands and develop large-scale plantations of crops after the con ict ( Management Unit 2020; UNHCR 2012). The grievances resulting from land con scation in the aftermath of con icts ultimately undermine peacebuilding and incentivize con ict recurrence (De Alban et al. 2019;Bhagwat et al. 2017;Woods 2015). The impacts of armed con ict on humans and the environment are likely similar across other con ict-affected regions in Myanmar, although different ecosystems, conditions that led to the onset of war, and the characteristics of the con ict itself need to be considered. In this context, areas in other con ict zones should also be investigated in the future to achieve higher accuracy of modeling results. Incorporating socio-economic surveys with high-resolution imagery might also help understand the land cover fully, and land uses dynamic during the study periods.
The ndings from this investigation further underscore the risk of forced repatriation of Rohingya refugees from Bangladesh. In addition to an absence of security, blatant disregard of basic human rights, and a lack of guarantee for the restoration of their housing, land, and property, the utter destruction of the ecosystem means sustainable livelihood solution is not likely upon return. Our post-con ict analysis highlighted that the Rakhine state's current situation is not yet equipped and cannot ensure the safe and voluntary return of refugees. Lack of fundamental human needs, livelihood opportunities, access to ecosystem services, among other factors, will make repatriation far from a durable solution (Naseh et al. 2018). Natural resources and environmental services underpin the livelihood of the majority of the population in the study area. If not properly managed, the combined pressure of con ict, environmental degradation, civil disorder, and the collapse of previously established livelihood systems will take a toll on both the civilians and the ecosystem. Moreover, post-con ict can be equally damaging to the environment if mass migration is reversed back to the point of origin, accompanied by resource over-exploitation necessary to rebuilt lives and revive livelihood opportunities. For example, in Afghanistan, the Middle East, and Northern Africa, coping strategies after livelihoods disruption due to war have led to widespread liquidation of the nation's natural resources such as forest cover loss, soil erosion, and water scarcity (United Nations Environment Programme 2009).
Therefore, in the reconstruction process, the environmental drivers and impacts of con ict must be managed. The fundamental needs such as health and livelihood are facilitated, and natural resources are used sustainably and distributed fairly to achieve stability and peace in the longer-term.
Undoubtedly, there will be no durable peace if ecosystem services that sustain livelihood are damaged, degraded, or destroyed. In addition to ending violence and persecution, failure to address resource management mechanisms will eventually result in con ict relapse.
The approach used in this study can help ll in the data gaps resulting from the restriction in conducting eld research in con ict zones. It is important to note that this methodology should only be used to complement eld-based research and not to replace them. The availability of ground-truth LULCC and precise environmental data would further improve the accuracy of the model. The use of mixed-method approach of quantitative GIS technology and qualitative led-collected personal narratives also might improve the results. Nevertheless, in the absence of eld data, this approach could be used across many domains and useful for producing granular data on environmental impacts. Future work may attempt to incorporate participatory approach using interviews with displace persons to validate ndings and provide a full sicario of the impact of armed con icts.

Conclusion
Rakhine's con ict is the greatest challenge faced by the military government (Tatmadaw) in recent decades and the civilian government since they took o ce in 2015. Concurrently, there are growing environmental challenges such as intense deforestation, land degradation, water quality deterioration, and climate change, biodiversity loss, and depletion of inland and coastal sheries. Land cover and land use in the country is changing rapidly and experiencing irreversible changes. Acute environmental health issues are on the rise, and ecosystem services are under an increased threat. On the other hand, natural resources and ecosystem services play a vital role in the national economy, but they are also critical to reduce Myanmar's vulnerability to climate change and natural disasters, poverty eradication, livelihood support, and peacebuilding. As mentioned above, Myanmar's long-lasting subnational con icts are often entwined with and fueled in part by abundant natural resources. While the con ict in Rakhine is not driven solely by economic interest, the abundant natural resource wealth and geostrategic location can be a signi cant factor. Although the key proximate causes of increased environmental pressure in Myanmar have been documented as large-scale development projects and shifting cultivation, the underlying drivers are deeply embedded within the country's socio-economic and political context. Land alteration during and after civil war are also recognized as the crucial drivers of environmental disasters, albeit they are still poorly understood phenomena in Myanmar. A critical analysis of armed con ict's environmental impacts provides vital information on ecosystem damage triggered by con icts and the mitigation strategies.
Our study examines the spatial and temporal pattern of environmental changes and the complex dynamics of land-cover-land-use shifts in war-torn Rakhine. By commissioning very-high-resolution-satellite-imageries and machine learning techniques, I set out to evaluate the capacity of very-highresolution-satellite data to quantify the magnitude of damage in con ict zones where the study context is notoriously di cult to assess. Remote sensing allowed for a comprehensive analysis of forest and land cover conversion in the Rohingya con ict zone in a 28590.724 ha study area in the Rakhine State of Myanmar from 2012 to 2019. Analysis reveals that the environment and ecosystem have undergone unprecedented change and have witnessed a dramatic forest cover decline. The region experienced a horrendous loss of residential area (nearly 100% wiped-out) regarding human life and livelihood loss. LULCC included an appalling expansion of burned areas and a barren and scrubland, resulting in a drastic reduction in all other forms of land cover, forest, wetland, human settlement, development, and agricultural lands. Forests predominantly covered the land, wetland ecosystems, planted/cultivated land before the con ict. However, all these land cover classes dropped off to virtually nil during the con ict, indicating a linear relationship between the environmental and human crisis and the con ict.
It can be concluded that over the eight-year study period, the environmental degradations and lost human lives, as well as livelihood distractions, are all devastatingly high, suggesting that the impact of the con ict is intense in Rakhine, given that all changes are related to burned areas. These results from Myanmar offer useful insights for other countries with ongoing armed con icts and civil war. The onset of armed con icts presents a signi cant risk to human lives, health, and livelihood; the environment becomes a silent casualty of con ict. Although the direct environmental impacts generated by chemicals and debris can be visible and well understood, the impacts on ecosystem services are di cult to determine. However, the disruption of critical natural assets and ecosystem services is equally important for reestablishing the community as they provide basic shelter, food and protection, and economic opportunities. Hence, these environmental impacts of wartime can be extremely persistent and widespread. The results from this research pinpoint that the environment and natural resources can be adversely affected throughout the con ict cycle, contributing and to the outbreak, being degraded by the violence, and undermining post-con ict reconstruction.
This research also provides a methodological framework for future studies investigating the impacts of armed-con icts using spatially and temporally explicit and quanti able methods through the integration of very-high-resolution satellite data, machine learning, and regional knowledge. Although costly, the mapping approach and the use of very high-resolution satellite data were relatively robust and provided very high accuracy and outcomes over a large geographic region without integrating multi-sensor data. The study also directs further attention to examining the severity of war's impact on the environment and overall human system. The end of hostilities in con ict zones might provide access to the study sites and an opportunity to corroborate ndings with eld data and further investigate underlying causes of the environmental crisis and further validate ndings from the satellite data in the future. Figure 2 Work ow for classifying environmental changes in the con ict zone in Rakhine Area change in land cover land use classes for three study periods Figure 4 2012 Pre-con ict, 2017during-con ict and 2019 Post-con ict land cover map for the study area classi ed into ve major land cover classes, including residential arear (pink), water/wetland (aqua), burned area (red), planted/cultivated (beige), barren/ scrubland (brown), development (blue), forest (green). Note: The designations employed and the presentation of the material on this map do not imply the expression of any opinion whatsoever on the part of Research Square concerning the legal status of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. This map has been provided by the authors.

Figure 5
True color composite map of the study area pre-, during-and post-con ict in Rakhine, Myanmar. Image A is from November 17, 2012, showing the area before the con ict. Image B is from October 28, 2012, showing the area during the con ict. Image C is from showing the area from November 26, 2019 after the con ict. Image D shows the location of the images in ROI. Source: https://www.intelligence-airbusds.com/geostore/. (Order number: SO200254-2-01) Note: The designations employed and the presentation of the material on this map do not imply the expression of any opinion whatsoever on the part of Research Square concerning the legal status of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. This map has been provided by the authors.

Figure 6
NDVI in the study region for the three periods. The colors represent values ranging between +1 to -1. Higher NDVI values indicate greener and healthier vegetation, while lower values represent stressed and depleted vegetation or barren land. Note: The designations employed and the presentation of the material on this map do not imply the expression of any opinion whatsoever on the part of Research Square concerning the legal status of any country, territory, city or area or of its authorities, or concerning the delimitation of its frontiers or boundaries. This map has been provided by the authors.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.