Using satellite data and machine learning to study conflict-induced environmental and socioeconomic destruction in data-poor conflict areas: The case of the Rakhine conflict

Models were developed using very high-resolution images (VHRI) to quantify LULCC. Nighttime light (NTL) images were supplemented to identify development disparities between the two regions. Pléiades 1A Satellite Sensor was used as the primary daytime satellite data used in the LULCC analysis. Pléiades 1A Satellite provides orthorectified VHRI data at 0.5-meter resolution with multi-spectral bands (AIRBUS 2020). We used the Suomi National Polar-Orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP/VIIRS) Day/Night Band (DNB) for nighttime data.

Despite providing abundant useful information, satellite data are extremely unstructured and make it difficult to produce reliable results at scale, even with intensive manual assessment (Jean et al 2016). The application of deep-learning and machine-learning methods for large-scale satellite data substantially improves object detection and classification (Feng et al 2018). The accuracy and effectiveness of these techniques can be achieved through supervised learning regimes with abundant good quality training samples (Hassan et al 2018). The training samples are usually gathered from the ground to acquire geofield photos and ground-truth data (Lecun et al 2015, Nomura and Edward 2018, Nomura et al 2019). In our setting, however, labelled data are lacking for Bangladesh and almost non-existent for Myanmar. The Myanmar population and housing census of Rakhine state does not include households in Maungdaw because the local population uses self-identified names that are not recognized by the government (Myanmar Government 2015). The Myanmar living conditions survey offers aggregate state-level data that are too coarse for our study purposes (Ministry of Planning and Finance 2018). Even in the circumstances where detailed household-level data are available, surveys usually contain datasets that are many orders of magnitude larger than the data typically useful for machine-learning classification (Jean et al 2016). Moreover, administering field visits to conflict effected area was logistically impossible due to travel restrictions and safety concerns.

To overcome this constraint, we extracted a comprehensive representative training sample based on the expert interpretation of a set of land cover in VHRI Pléiades satellite data for each period. In using VHRI for image classification, this approach provides higher accuracy results with relatively little error compared to collecting training and validation data in Google Earth imagery. Following the experts' regional knowledge of the environment and socioeconomic conditions as well as land uses and land cover (LULC) changes in the study area, we physically delineated reference data through visual interpretation, developing each datasets for the three study periods. As our study aims to compare the status of the landscape and socioeconomic changes between the two study areas at the local level rather than estimating poverty at the household level, the use of proxy measurements such as household expenditure or assets was not necessary. We generated and digitized a total of 4,663 to 4,682 training and testing reference polygons for each year dispersed throughout the area of interest covering the entire satellite image.

Land cover classifications were identified according to the historical land cover maps of Rakhine provided by the Myanmar Forest Department (2018) as well as the land cover classification and analysis map of Rakhine state produced by the United Nations Operational Satellite Applications Programme (OCHA 2015). The map of Bangladesh provides the land cover classification of the entire country by region. The maps were developed using Landsat 5 TM and Landsat 7 ETM images at a 30 m pixel resolution. The classification used the national land representation system of Bangladesh developed by the Food and Agriculture Organization (Jalal et al 2016) and covered 24 land cover classes. The Rakhine state land cover offers land cover classification produced from Landsat 8 multi-spectral imagery at a 30-m pixel resolution. The map categorized five main classes namely forest, mangrove, cropland (paddy field), barren soil, and vegetation. As the map uses are broadly defined classes, we supplemented these with a countrywide land cover map (SERVIR 2018).

In this study, the measurement of economic and environmental well-being requires micro-level detailed classification of land cover classes and features. Thus, we classified a total of 12 classes under seven broadly defined categories: (1) residential building, (2) infrastructure, (3) agriculture, (4) wetland, (5) water, (6) barren or scrubland, and (7) forest. We then subclassified residential area into three classes based on roof type: aluminum, concrete, and thatched. This subclassification was done based on the common type and structure of houses in both Myanmar (Oo et al 2003) and Bangladesh (Hasan 2000). We also subclassified the infrastructure into transportation infrastructures featuring road networks and energy infrastructures. Energy infrastructure was identified based on the appearance of solar panels, wind turbines and grids. Next, the wetlands were subdivided into three classes—fish pond, mangrove, and swamp—based on the regional characteristics (Hredoy et al 2018, Zöckler et al 2018). These reference polygons and training samples were applied as input variables in the calibration of the random forest (RF) algorithm. R-studio programming software was utilized to create a binary column in the reference dataset to split polygons into the 50% for training and 50% for testing the model.

Primary data from the Pléiades 1A Satellite Sensor were used to predict environmental and human welfare through the quantification of LULC changes. Two years of data were captured for both areas using similar dates. To conduct comparison among pre- and post-conflict intervals in the region, the satellite data from two dates (t0 = November 11, 2012 and t1 = November 28, 2019) were used for our study areas in Maungdaw and Teknaf. The VHRI data were primarily used to track and classify LULCC in the conflict and migration zones based on the RF machine learning model. The study dates were selected according to the times at which the conflict and mass migration are at their highest intensity and when cloud-free VHRI images were available. Monsoon climate patterns in Myanmar makes obtaining cloud-free satellite imagery often challenging. ArcGIS Pro was used to pre-processed and mosaicked the images before classification. Firstly, apparent reflectance function in ArcGIS Pro was used to adjust surface reflectance or brightness based on the scene illumination and sensor-gain settling. This calibration allows less variation between scenes from different dates before color balancing, mosaicking and classification. Table 2 below details the satellite data description.

To compute these outcomes, Google Earth Engine (GEE), a cloud-based computing platform, was used to implement the RF model. Preprocessing and change detection of large-scale high-resolution satellite imagery can be very time-consuming and high computational capacity capable of process large volume to data. GEE is a cloud-based alternative that offers increased efficiency of planetary-scale geospatial analysis (Gorelick et al 2017, Shelestov et al 2017, Sidhu et al 2018, Phan et al 2020). Figure 2 presents the workflow of this project. First, Google's Cloud storage was used to store the satellite images and training data and integrated it into GEE assets through Python's application programming interface. Next, we derived training and testing data in GEE by updating the training parameters and testing data by assigning datasets and matching bands. Lastly, we included ALOS Global Digital Surface Model (DSM) developed by Japan Aerospace Exploration Agency (JAXA) to add new bands for additional exploratory variables, including the digital elevation model and slope. The maps consist of Digital Elevation Model (DEM) or Digital Surface Model (DSM) that can represent land terrains with five meters in spatial resolution (JAXA 2017).

Zoom In Zoom Out Reset image size

The normalized difference vegetation index (NDVI) was calculated with Pléiades 1A images by applying the following formula:

$$\begin{eqnarray}&&NDVI=\displaystyle \frac{{\rm{NIR}}-{\rm{Red}}}{{\rm{NIR}}+{\rm{Red}}}\end{eqnarray} \tag{ 1 }$$

Where NIR is Near Infrared band and, and Red is the red band.

The classification of LULCC was conducted using VHRI images using the RF classification model. RF classification is a non-parametric machine learning algorithm widely applied in remote sensing and classification modelling (Gislason et al 2006, Cutler et al 2007, Horning 2010, Sesnie et al 2010, Rodriguez-Galiano et al 2012, Bricher et al 2013, Belgiu and Dra 2016, Ming et al 2016). We tested two widely used supervised classification methods in remote sensing research, RF and Support Vector Machine (SVM) to compare their overall accuracies. In our study, the RF classifier obtained higher classification accuracy with above 80% for all the study periods whereas SVM's accuracy ranged only from 32% to 45%.

The two crucial parameters of RF are the creation of an ensemble of trees, each assigning a 'vote,' and the number of variables experimented at each split to choose the best classification method. Most of the votes from the assemblages of the tree constructed in RF identify the class assignment of the pixel, and the results obtained a large number of trees are internally aggregated (Berhane et al 2018). Before the classification step, hyperparameter tuning of the model was conducted to update the bands and determine the best performing hyperparameter values. The execution of RF model requires the specification of several parameters. Hence, each RF tree was created by training each tree in the forest (ntree) with the number of input predictor-variables (mtry). The variables are randomly selected at each split from the training dataset (Aung et al 2020). We applied 200 decision tree and 10 minimum leaf population at each time.

Next, building on the knowledge gained from this image classification task, we calculated NTL intensity corresponding to daytime imageries used in both study regions. We extracted monthly NPP/VIIRS NTL images and calculated annual NTL intensity for the years 2012 to 2018 by averaging monthly composites. We calculated the state-wise sum of NTL intensity for Chittagong and Rakhine rather than precisely predicting Teknaf and Maungdaw, as the study areas are too small to produce any meaningful NTL data. NTLs are a noisy yet globally consistent and globally available proxy for welfare measurement (Xie and Weng 2016, Jiang et al 2017, Li et al 2020).

The accuracy assessment was first evaluated from RF model during the bootstrapping process and produced by GEE (Belgiu and Dra 2016). However, following the best practices put forward by Olofsson et al (2014), we supplemented a stratified random sampling design to conduct detailed accuracy assessment of the model. The assessment of accuracy of the model and the resulting maps are necessary to evaluates the errors of the classification and the uncertainty of the information generated (Mellor et al 2013, Sharma et al 2017). The sample size required in the assessment was computed based on the following formula:

$$\begin{eqnarray}&&n=\displaystyle \frac{{\left({\sum }^{}{W}_{i}{S}_{i}\right)}^{2}}{{\left[S\left(\hat{O}\right)\right]}^{2}+\left(\displaystyle \frac{1}{N}\right){\sum }^{}{W}_{i}{S}_{i}^{2}}\approx {\left(\displaystyle \frac{{\sum }^{}{W}_{i}{S}_{i}}{S\left(\hat{O}\right)}\right)}^{2}\end{eqnarray} \tag{ 2 }$$

where n is the number of units; $S\left(\hat{O}\right)$ is the standard error of the estimated overall accuracy; W_i is the mapped proportion of the area of class i; and S_i is the standard deviation of i, which is found using the equation S_i = $\sqrt{{U}_{i}\left(1-{U}_{i}\right)}.$ We identified a target standard error for an overall accuracy as 0.01. Using the proportional approach, a sample size of 50–100 was allocated to the smaller classes, and the remaining samples were proportionately assigned for each change strata. The predicted variances were calculated based on the sample size allocation.

The objective of an accuracy assessment is to assess the ability of the model to detect and delineate changes within a study area during the study period. The accuracy results internally generated from the RF algorithm showed satisfactory accuracy for Bangladesh with over 85% in 2012 and 82% in 2019. The overall accuracy is slightly lower for Rakhine with 71% in 2012 and 76% for 2019. The results are summarized in the supplementary table 4 (available online at stacks.iop.org/ERC/3/025005/mmedia). Producer's accuracy (omission error) refers to how often are true features on the ground accurately shown on the result map or the probability that a specific land cover of an area on the ground is classified accordingly. User's accuracy (Commission error) refers to as reliability or how often the class on the map will truly present on the ground. The producer's and user's accuracies for Planted/cultivated were the highest for all years. The second highest accuracy was obtained by the fishponds. The aluminum roof categories also received relatively high accuracy. On the other hand, mangrove and forest had the lowest accuracy for all the years. The lower accuracy can be due to the model's confusion to differentiate among forested land and mangroves. Therefore, scrubland and barren also received relatively lower results. Additionally, accuracy assessment results in area propotions and sample count using Olofsson et al (2014) method are presented in supplementary table 1 to 3. The classification's overall accuracy was 93.90% for 2012 and 94.26% for 2019 in Bangladesh. For Myanmar, the overall accuracy was 91.60% for 2012 and 95.05% for 2019. The confidence interval for all years was 0.95. The supplementary material delineates the classification accuracy validation results and error matrices of the classified maps.

Our satellite imagery analysis through machine-learning model is strongly predictive of local-level human welfare and environmental conditions in both conflict-affected regions and migration destinations. The comparison results between Maungdaw and Cox's Bazar show striking differences in the level of development, environmental changes, and livelihood conditions (tables 2 and 3; figure 3 and 4). The classification maps of the study area from two time intervals were investigated, including the pre-conflict (2012) and post-conflict (2019) periods. This way, the environmental and welfare conditions in both regions were identified, and the destruction due to conflict and mass migration was predicted.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Type of building structure (aluminum, concrete, or thatched), fishpond, and planted/cultivated measures were used to indicate socioeconomic conditions. The building type category is reflected by structure and texture variables and can indicate the physical living condition of households. The development level can be determined by examining the comprehensive landscape features, such as the existence/non-existence of infrastructure (road networks and energy). The road network variable describes the accessibility and ease of access necessities for households. These outcomes were complemented with NTL data to quantify and compare the level of welfare in the two regions.

The environmental changes were measured by LULC changes (mangrove, swamp, barren/scrubland, water, and forest). The burned area category was added to identify the direct impact of conflict in Rakhine. Whereas previous socioeconomic research based on remote sensing focused on income or wealth estimates across the population, the central objective of this study is to compute and compare living condition and environmental degradation between the two regions and examine the impact of armed conflict and mass migration.

The machine-learning classification results showed entirely different patterns of residential regions between the whole study area. In general, the results from the type of building suggest that the socioeconomic conditions in Teknaf were incomparably higher than Maungdaw, even before the conflict in Rakhine state. The results from 2012 show that in Teknaf, there is an equal distribution of roof types among all the buildings. Based on the regional characteristics, roofs classified as aluminum or concrete are considered to be associated with a lower level of poverty than thatched roof types (Hall et al 2001, Tarmizi et al 2014, Engstrom et al 2017).

In total, 55.29% of the land was covered with buildings (aluminum + concrete + thatched). There was a 6.02% increase in the residential area between the two study periods. This increment could be due to the expansion of camps to host the exodus of Rohingya refugees during the two periods. (Hassan et al 2018) reported on the significant expansion of refugee community in Teknaf, which increased from 175 to 1530 hectares from 2016 to 2017 (net growth of 774%). It is important to note that high roof coverage indicates population density, which is a typical characteristic of slum areas (Kuffer et al 2016). For Maungdaw, 97.53% of the houses had thatched roofs in 2012, indicating a much lower level of poverty than in Teknaf. Aluminum and concrete roof types represented only 1.91% and 0.56%, respectively. More staggering results were observed in 2019, as the total residential area declined to just 7.97% from 34.03%. Thatched roofs remained the primary roof type, covering 98.08% of the remaining structures. This result is also reflected in the expansion of burned areas in 2019 (2.7%), even in the post-conflict period. Although extensive land recovery can happen within two years after a conflict, the damage could still be observed.

Another socioeconomic variable in our study was the livelihood situation and income sources as measured by the proportion of agricultural lands and fishponds, as the populations in both Maungdaw and Cox's Bazar are mostly dependent on fishing and agriculture as the main sources of income (Bhattacharya and Khan 2016, Gupta 2016). Both regions experienced a decline in fishpond area in 2019. For agricultural lands, during the pre-conflict period, Maungdaw had a higher percentage of planted/cultivated areas compared to Teknaf. However, in 2019, as a result of burning and land abandonment during the conflict (Witmer 2008), Maungdaw experienced a considerable decrease in agricultural lands. In contrast, agricultural land in Teknaf expanded. This expansion in agricultural land may have resulted from the increase in the refugee population and thus the use of land for growing food. The same situation was observed in the surrounding areas of the internally displaced person camp in Darfur (Lang et al 2010, Spröhnle et al 2016).

Next, the accessibility variable, road networks, also revealed substantial evidence for development. Generally, an increase in road networks signals steady regional development and the local populations' increased accessibility to necessary facilities such as health care and education services (Angeles et al 2009, Zhao et al 2019). Only 2.36% of the total land area in Maungdaw was occupied by roads. In terms of measuring access to electricity, the energy variable was included in addition to NTL data. The model showed a complete absence of energy infrastructure in Maungdaw. This was due to the negligible number of objects detected during the RF training process. In Bangladesh, there was 1.043% of land area was covered by energy infrastructure in 2012, and the number increased to 2.680% in 2019. This expansion was the result of the installment of an extensive solar power plants in 2018 in the Teknaf region (Ministry of Power, Energy, &. Mineral Resources 2015). There was a small area (45 ha) of burned area in 2012 in Maungdaw. This could have been somewhat attributed to the scattered burning of houses during the first wave of conflict in 2012, as well as due to winter slash-and-burn practices.

In addition to socioeconomic variables, land cover features measuring environmental conditions can be vital indicators for human well-being, especially in rural areas, where a predominantly large number of the population relies on ecosystem services for their livelihoods (Tatem et al 2013, Yeh et al 2020). The linkage between land cover pattern and socio-economic condition is particularly pronounced in conflict-affected contexts (Zúñiga-upegui et al 2019). In Maungdaw, forests covered 6368 ha (over 19% of total land cover) in 2012. In 2019, the proportion of forest area was reduced to 5108 ha (15.01%), showing an overall deforestation rate of 20.05%. Meanwhile, Bangladesh also experienced deforestation, with a total forest loss of 12.98%. Deforestation is a typical damage pattern observed in conflict-affected regions and migration destinations globally (Stevens et al 2011, Gorsevski et al 2012, Hagenlocher et al 2012, Gray and Bilsborrow 2013, Nackoney et al 2014, Butsic et al 2015, Eklund et al 2017, Salerno et al 2017, Brito et al 2018, Hassan et al 2018).

A similar situation was seen in the mangrove forests, which were reduced by over 23% in the study area in Maungdaw by 2019. The mangrove degradation in the migration destination, Teknaf, was more pronounced at 36.84%. Additionally, there was a sharp increase (604.8%) in barren/scrublands in Maungdaw in 2019. There was also a notable increase in swamp area. These increases were driven by large-scale burning, agricultural land abandonment, and the destruction of aquaculture areas, which is common in conflict settings (Witmer 2008). This result is also reflected in an excessive loss of agricultural lands. In contrast, Teknaf underwent a sharp decline in barren land. It is likely that some previously barren land was converted into camps to host refugees (see figure 4 above). There was also a slight increase in swamplands in Teknaf. The changes in the water areas of both regions were minor.

The NDVI maps is in accordance with the evidence of environmental degradation in the region. The NDVI maps indicated that NDVI values substantially decreased during the conflict periods (figure 5). For Maungdaw, the mean NDVI value for 2012 was 0.625 (SD = 0.196). The mean NDVI value reduced to 0.446 (SD = 0.281) in 2019. For Teknaf, the mean NDVI value was 0.436 (SD = 0.224) in 2012 and 0.591 in 2019 (SD = 0.253). The results indicate that the vegetation health and greenness of Maungdaw were much better before the conflict and declined to values lower than those of Teknaf during the conflict years.

Zoom In Zoom Out Reset image size

Figures 6 and 7 show the results of the NTL analysis. The figures demonstrate distinctive differences in the electrification rates between the two regions. The NTL intensity in Teknaf was consistently higher than in Maungdaw during the study periods (see figure 6). The increase in NTL intensity in the proximity of the refugee camp indicates access to electricity in the camp. Figure 7 shows the graphical representation of NTL intensity. In addition to a higher electrification rate, the acceleration of NTL in Bangladesh was visibly higher than in Rakhine. This elevation in NTL intensity could be contributed to the development of the nation's largest solar power plant in Teknaf (see figure 8).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size