Our region of interest (ROI) consists of three countries: Myanmar, Thailand, and Cambodia. Vietnam was originally included in the workflow, but due to the low quality and quantity of observations, the results were excluded in the study. The ROI resides within a tropical monsoon and rainforest climatic zone with temperatures above 25°C throughout the year (Peel et al., 2007). Mangrove forests are composed of trees and shrubs that are adapted to the saline and brackish conditions of the ROI coastline. They are taxonomically diverse plant species and occupy 42% of each coastline in the ROI (Bunting et al., 2022). Mangroves in Thailand are found consistently along muddy tidal flats or at the base of river mouths along the southern and eastern coasts, with a two-story forest structure (Pumijumnong, 2014), and at higher densities around the Gulf of Thailand and Andaman Sea. The lower story of mangroves in Thailand grow more than 20 m and are dominated by species from the Rhizophora, Herritiera, and Xylocarpus genera. The upper story mangrove species of Thailand include the Bruguiera and the Ceriops genera, with some of these species like the Bruguiera gymnorrhiza growing more than 40 m above the forest and with trunks as thick as 2 m (Aksornkoae, 2012; Pumijumnong, 2014). Myanmar hosts an array of mangrove species just as numerous and diverse from the Rhizophora, Avicennia, Bruguiera, Ceriops, and Xylocarpus (Zöckler and Aung, 2019) genera and are primarily found in the Rakhine, Ayeyarwady, and Tanintharyi divisions (Zöckler and Aung, 2019). Mangrove forests in Cambodia are found primarily in four provinces, Koh Kong, Sihanoukville, Kampot, and Kep (Nop et al., 2017; Kozhikkodan Veettil and Quang, 2019). The most found species in Cambodia are from the Rhizophora, Nypa, Bruguiera, Ceriops, Lumnitzera, Heritiera, Xylocarpus, Hibiscus, Phoenix, and Acrostichum genera.

The methodology (Figure 1) used to produce the historical extent maps was divided into six steps – delineation of the ROI, evaluation and selection of Landsat 1-2 MSS (LMSS) Collection 1 Tier 2 DN, processing and correction of LMSS, generation of a 1970s mangrove map, assessment of accuracy, and a comparison of the baseline to Global Mangrove Watch (GMW) mangrove extents (Bunting et al., 2022). All processing and analyses were carried out using the Google Earth Engine Platform, ArcGIS Pro, and R. Figure 2 displays the coastal mask used to delineate the region of interest, pixels assigned with a value of zero due to no Landsat observations, the Landsat scene WR-1 path and row, and the 1970s mangrove extent. The final LMSS composite and the 1970s mangrove classification can also be referred to in Figures 1, 3.

Step one in our workflow included data filtering and selection followed by a series of pre-processing steps. To delineate coastal areas and subset the LMSS data needed for processing, a coastal mask was generated to include all potential mangrove areas. The mask included: the known extent of more current mangroves and coastal wetlands based on the global Wetland Extent Map (Mcowen et al., 2017) and the GMW dataset (Bunting et al., 2022), areas within 10 km of the shoreline, based on the global shorelines data (Sayre et al., 2019) and areas lower than 20 m elevations based on the Shuttle Radar Topography Mission (SRTM) DEM (Farr et al., 2007; Yang et al., 2011).

Given the unavailability of Tier 1 Landsat data, we utilized LMSS Tier 2 data for our analysis. Scenes with this level of processing have a lower radiometric and positional quality, but due to the lack of available scenes, this collection was selected. The coastal mask was used to select scenes from LMSS Tier 2, which were scaled and calibrated to at-sensor radiance. Only systematic terrain (L1GT) and systematic (L1GS) processing were applied to this collection due to insufficient ground control, excessive cloud cover, and geolocation issues (U.S. Geological Survey Department, 2018). This collection has a resampled spatial resolution of 60 m and a spectral range of 0.5–1.1µ including the Green, Red, NIR-1, and NIR-2 channels, and corresponds to WRS-1 (Wulder et al., 2022; U.S. Geological Survey Department, 2018). All scenes available from 1972 to 1977 were selected using the coastal mask as a spatial filter, which was then followed by a series of exclusionary steps. Scenes that did not have all of the five bands present which are the Red (B4), Green (B5), Near Infrared 1 (B6), Near Infrared 2 (B7), and the quality assurance bitmask (QA_Pixel) bands were excluded. Additionally, scenes were excluded if they had L1GS processing, exceeded a spatial displacement greater than ~24m or a root mean square value of 0.4, and cloud cover greater than 30%; effectively reducing imagery from 3,153 to 689 (see Figure 4). Images were then manually excluded if they had excessive cloud cover over mangrove areas, erroneously saturated pixels, or abnormal image artifacts, further reducing the collection to 371 images (see Figure 5). These issues were related to excessive detector striping, transcription artifacts, abnormal saturation, memory effect, and scan mirror pulse errors (U.S. Geological Survey Department, 2018; Vogeler et al., 2018).

Following the manual inspection and removal of Landsat scenes from Landsat MSS over the study area, the next step was to mask out cloud cover and correct the imagery to top-of-atmosphere reflectance. We did this by using an automated cloud and cloud shadow detection and masking algorithm proposed by Braaten et al. (2015), called the Landsat Multispectral Scanner System clear-view-mask or MSScvm. This algorithm is an already established automated approach that identifies and masks out clouds based on green band brightness and the normalized difference between the green and red bands. It further identifies and masks cloud shadows using near infrared band darkness and cloud projection. Along with this cloud detection, it also corrects for topography-induced illumination and water identification using a digital elevation model (Braaten et al., 2015). Then on a per image basis, MSScvm was applied to each image effectively removing the majority of cloud cover and cloud shadow. Following that, we created a mosaic for subsequent classification. As shown in Figures 4, 5, the amount of imagery is highly variable. This study selected the years 1972 to 1977 due to these years having the highest and most semi-consistent image count with almost wall to wall coverage (see coastal mask in the Figure 1 study area) during the 1970s. Once the images were identified, the cloud free mosaics were generated per calendar year from 1972 to 1977 on a per pixel basis using the medoid, a more robust version of the median. Although these studies are more effective with seasonal composites, this was not feasible due to low image counts. The medoid represents the point with the minimal summed distance to all points in a data set. It also takes into consideration the multiple dimensions in the selection relevant to the different bands of a scene found within a year (Flood, 2013). The medoid was calculated by finding the difference between the median and the corresponding median of each band squared followed by finding the sum of the powered differences across bands per image for all observations. These annual medoid mosaics were much more sensitive to extreme outliers resulting in a more consistent representation of the study area. This approach was also selected to account for the inconsistent number of Landsat observations per year. The yearly medoid image collection was then used to produce a single five-year composite over the study period using the median across all medoid observations (see Figure 3 for composite results and Figure 1 for workflow).

Following the steps used to create a single mosaic of LMSS from 1972–1977, a series of indices were calculated that are ideal for mapping mangrove extents (Yancho et al., 2020). The indices that were calculated include the simple ratio (SR), normalized difference vegetation index (NDVI), the normalized difference water index (NDWI), the combined mangrove recognition index (CMRI), the soil-adjusted vegetation index (SAVI), the optimized soil-adjusted vegetation index (OSAVI), and the greenness chlorophyll vegetation index (GCVI) which serves as a proxy for chlorophyll content and proved to be useful in some mangrove mapping studies (Jordan, 1969; Tarpley et al., 1984; Huete, 1988; Gao, 1996; Gupta et al., 2018; Yancho et al., 2020; Rondeaux et al., 1996; Gitelson et al., 2003; Chamberlain et al., 2021). In addition to the vegetation indices, slope and elevation were also incorporated which was extracted from JAXA’s Land Observing Satellite (ALOS) (Takaku et al., 2014; Takaku et al., 2020). The predictors used for the random forest classifier can be referred to in Table 1 in addition to the original bands used in the LMSS collection.

Before beginning the process of collecting model calibration and validation samples, masks were applied to the coastal region of interest (see Figure 1) using the ALOS DSM (Takaku et al., 2014; Takaku et al., 2020) and NDVI bands. These datasets were used to exclude areas that were not vegetation using NDVI pixels less than 0.05 and areas of elevation greater than 40 m. This masking excluded any area that did not have an NDVI values associated to dense vegetation, such as urban, water, and bare ground areas in addition to higher elevation areas. Given this level of masking, this study specifically focused on dense assemblages of mangrove forests and excluded only the most fragmented mangroves with low NDVI values. Training polygons were collected using the final composite by digitizing areas representing the most homogenous mangrove and non-mangrove vegetation pixels. Due to the lack of reference data from the 1970s, the composite itself was used as a reference to digitize the training samples. In total, 1,134 points were selected for the mangrove (n=283) and non-mangrove (n=851) land cover classes. Using the Landsat MSS composite (see Figure 3) an area was designated as mangrove if it was found within the coastal mask region, was found along the coastline, found along tributaries closest to the coastline, had a dense texture, and was a patch with high and consistent NDVI values. The non-mangrove training areas were selected based on whether it was found in water, areas at greater distances from the coast, and in groupings of pixels that were not homogenous or did not have a high NDVI value. These characteristics were chosen given that the minimum mapping unit (MMU) was around 0.16 ha and that we had a goal of mapping an assortment of mangrove trees and not individual mangrove trees below the MMU.

Following the preparation of predictors and collection of training samples, a random forest algorithm was used to predict the distribution of mangrove and non-mangrove areas for the entire study area using the LMSS mosaic. Random forest is an approach that uses non-parametric classification and decision trees in addition to classification and regression trees (Breiman, 2001). The hierarchy of this classifier is composed of a root node, inclusion of predictor samples, node separator with relevant decision rules, and the end of the leaf node with the desired classes – or in our case the probability of belonging to the mangrove class. Random forest was also chosen because models in other studies resulted in higher accuracies in comparison to other decision tree classifiers (Breiman, 2001; Pal, 2003; Ghimire et al., 2012; Belgiu and Drăguţ, 2016). The predictors that were selected for the model used in this study included the previously prepared indices or elevation parameters; SR, NDVI, NDWI, CMRI, SAVI, OSAVI, GCVI, and DSM. The random forest model we selected utilized a total of 100 trees sampled at random for every 5^th predictor, a minimum leaf population of 1, a bag fraction of 0.5 per tree, no limit on the maximum number of leaf nodes in each tree, a randomization seed value of 0, with the output mode set to a probability output. Following this classification of continuous mangrove probability, a series of post classification steps were used to remove noise and to threshold the classification’s bimodal distribution. First the classification was automatically thresholded using a gray level histogram detection method proposed by Otsu (1979) to divide the layer into two distinct classes, mangrove, and non-mangrove. Once the classification was automatically converted into a binary classification of mangrove and non-mangrove areas, the classification was cleaned up to remove noise using a majority filter. The majority filter was applied using a 3 by 3 kernel majority filter where a given pixel would be changed if most of the cells within a neighborhood were contiguous. Following the majority filter, a series of dilation and erosion techniques were used to generalize the zones using an evaluation of immediate orthogonal and diagonal neighbors for a given mangrove or non-mangrove region (Serra, 1982). The order of sorting priority was based on the size of the mangrove and non-mangrove zones when performing the classification smoothing. The classification was manually cleaned up to eliminate additional salt and pepper areas, areas that had visible errors introduced by excessive cloud cover or had any additional noise. An example of this would be any errors of commission found farther from the coastline where brackish waters are less concentrated.

To assess the classification, we followed the “good practices” proposed by Olofsson et al. (2013) and Olofsson et al. (2014) using a random stratification approach over the study area. It is important to note that the area proportion for the mangrove and non-mangrove class were 16% and 84% of the total study area and the sample allocation was based on the smaller area proportion class (mangrove). Furthermore, we anticipated an accuracy of 70% and a proportion of 20% for the mangrove class, and a target standard error of 1%. This leveraged a set of 300 samples for the non-mangrove class and a set of 67 samples for the mangrove class. However, a total of 28 were impossible to verify visually, therefore our final dataset resulted in 339 samples: 54 to mangrove and 285 to non-mangrove. The verification process took place in the Collect Earth platform (Bey et al., 2016; Saah et al., 2019), each sample was inspected using both the original LMSS mosaic, Google Earth Imagery (Lisle, 2006), and the classification overlaid as a reference. Sample points were validated as mangrove when 50% of a 30 m x 30 m square centered about a sample point had a higher NDVI reflectance (manifested as a dark red color), were adjacent to the coastline, at the intersection of a river outlet and the ocean, or were greater than the MMU of ~0.16 ha. The non-mangrove class was labeled if a given sample point was found in open water, bare ground with some vegetation cover, water with some vegetation, heavily fragmented vegetation, and vegetation that was not immediately adjacent to the coastline or river outlets for greater than 50% of a 30 m x 30 m area. Then the attribute table of the validation point was updated by extracting the value found in the classification at each point location. This was used to calculate the error matrices, overall, producer’s, user’s accuracy, and Kappa Coefficient.

Once the classification and accuracy assessment was completed, we analyzed the extent of mangrove ecosystems for our ROI within the context of each country’s unique circumstances, GMW extents, and other external estimates. In order to measure change in mangrove extent between the 1970s and 2020, we compared our 1970s mangrove cover map to maps created by GMW for 1996, 2007, 2010, 2016, and 2020 (Bunting et al., 2022). The GMW maps were reprojected and resampled to match the spatial resolution of the LMSS classification results. Then, the layer was rasterized. Lastly, the GMW raster was masked to exclude all areas that overlapped with no Landsat observations available in the final LMSS classification product.

A total of 3,153 images were identified during the initial data exploration phase before additional filters were applied. This study identified 371 images suitable for classification that met spatial offset and cloud cover filters, which reduced the available imagery by 88%. The study had an average of 63 ± 38 images per year but with significant interannual variability. The highest image availability occurred in 1973 (n = 109) and 1976 (n = 103). The year with the lowest image count was 1974 with 9 images. Following the filtering and scene selection phase, the LMSS imagery was corrected to top-of-atmosphere, cloud and cloud shadow masked, and then composited into a mosaic over the study period resulting in a cloud free medoid mosaic of LMSS. This approach likely has high amounts of variability and uncertainty due to the lack of observations & the inability to map mangrove cover seasonally. Additionally, the coarser spatial, spectral, and radiometric resolution of MSS constrains the capacity for mapping mangroves during the 1970s. This work would greatly benefit from additional efforts to map mangroves on a seasonal basis over a longer time period consistently with harmonization across all of Landsat’s sensors as done by other studies (Braaten et al., 2015; Zhu et al., 2018; H. Nguyen et al., 2019; Yan and Roy, 2021). Following the classification, the final mangrove data product was assessed for accuracy, but with other constraints.

The overall accuracy and kappa coefficient was 95% and 0.82 respectively. Producer’s and User’s accuracy were 98% and 96% for the non-mangrove class and 80% and 90% for the mangrove class (Table 2). These accuracy results indicated that the model was less likely to identify real mangrove features on the ground in comparison to the mapped mangrove feature in the map. The higher reliability, or User’s accuracy for the mangrove class indicated that map users were more likely to find the mangrove areas identified in the map on the ground. Some of these uncertainties were likely related to the quality of the mosaic and how consistent a given pixel may be. Take for example the issue of low image counts for 1974 in this study and the fact that there were regions with no Landsat observations.

Our 1970s map of mangrove extent identified 15,420 km² across our study area (Table 3). Myanmar had the greatest extent of mangroves (9,272 km²), followed by Thailand (5,407 km²) and Cambodia (742 km²). When comparing our new 1970s baseline data to existing GMW maps of mangrove extent, we found a sharp decline in mangrove area between the 1970s and 1996. Across the study region, mangrove area declined by 47% (corresponding to a loss of 8,239 km²) during this period (Table 3). Loss rates were highest in Thailand (58%) and lowest in Cambodia (14%). In Thailand, declines in mangrove areas were especially pronounced around Bangkok and along the eastern portion of the Gulf of Thailand (Figure 6). In contrast, areas with the greatest occurrence of persistence and gain were found in the Mu Ko Phayam National Park along the coastline up to the Mu Ko Ra-Ko Phra Thong National Park. Myanmar experienced a 42% decline in mangrove area between the 1970s and 1996. Here, loss was greatest around the Ayeyarwady Delta (Figure 7). In Cambodia, most losses were concentrated along the Koh Kong coastline in the northwestern portion of the country (Figure 8). Although Cambodia experienced overall declines in mangrove extent between 1970s and 1996, there were some southwest gains in the Bay of Kampong Som in the Botum Sakor District of Koh Kong Province (Figure 8).

Firstly, the GMW estimates were masked using the areas labeled as 0 due to a lack of Landsat observations in the LMSS mosaic. The inventory of extent of different time points up to 2020 show that rates of mangrove loss across our study region declined dramatically after 1996 (Table 3; Figure 9). Overall, mangrove extent declined by 7% between 1996 and 2007 and these rates were similar across the three countries. Mangrove extent then showed little change between 2007 and 2016. Between 2016 and 2020, mangrove extent increased in all three countries; rates of mangrove expansion ranged from 7% (Cambodia) to 18% (Thailand). These results further extend our understanding of mangrove change before the 1990s, a period of data scarcity when it comes to mangrove data products.

In this study, we provided a semi-automatic approach to delineating the extent of mangrove area using LMSS data from the 1970s. We leverage the earliest available Landsat imagery using MSS and compared it to more recent mangrove extent maps to understand historical mangrove change. However, some limitations remain, for instance none of the LMSS data from this period has been processed to the highest level of terrain and precision (Braaten et al., 2015; Roy et al., 2016). The issues that affected the quality of the collection, a sensor that makes up a significant portion of the early Landsat record (Yan and Roy, 2021), were related to a variety of anomalous errors such as memory, effect, scan correlated shifts, and scan mirror pulse¹ . In addition to sensor-related issues, there is also the difficulty of conducting a remote sensing investigation in the case study countries, one of the planet’s cloudiest regions (Kontgis et al., 2015; Li et al., 2018; Oliphant et al., 2019). Because of these challenges, this study excluded imagery that had a spatial offset greater than 24 m, cloud cover of less than 30%, imagery with oversaturated pixels, or excessive striping. It should also be noted that the number of images found per year was not consistently the same (Figures 4, 5) or of an ideal quantity to do annual image classifications and change detection. As a result, we had to composite multiple years’ worth of data to cover the entire study area. These challenges will make it difficult to expand this approach to other countries outside of our study area.

Following this phase, careful measures were taken to identify the LMSS scenes that would be used in this study. The images then had to be preprocessed to TOA and cloud masked using the MSScvm (Braaten et al., 2015) algorithm. This algorithm was developed by researchers to overcome the challenge of automatically detecting and masking clouds from MSS (Braaten et al., 2015), but with some limitations. This algorithm was designed to work with temperate ecosystems, but due to the lack of automatic algorithms for MSS cloud detection, MSScvm was selected for pre-processing procedures. Many measures were taken to identify the best approach to establishing a new baseline, but these limitations must be considered.

With our new 1970s baseline map of mangrove extent in our case study countries, we identified a sharp decline in mangrove extent primarily for Myanmar and Thailand between the 1970s and 1996. However, after 1996, mangrove loss rates declined dramatically, and mangrove extent has been relatively stable since the mid-2000s according to this assessment. We do believe that an additional effort must be done to map the extents more consistently and harmoniously across sensors over a dense time series. But that was not possible at this time because of the lack of MSS scenes of sufficient quality. The proximate and underlying drivers of gains and losses for our study area are complex due to the history of political and economic instability from situations like the reign of the Communist Party of Kampuchea (CPK) over the course of 1975 to 1979 in Cambodia (Mosyakov, 2000) or the lack of electricity and subsequent need for mangrove firewood (from the Ayeyarwady delta) in the Yangon province of Myanmar. Furthermore, studies have indicated that some of the Myanmar and Cambodia mangrove losses may have been attributed to Thailand’s ban on logging in 1989 (Brown et al., 2001) proving that conservation measures can have unintended deforestation consequences (Brown et al., 2001; Pumijumnong, 2014; Lim et al., 2017). Meanwhile, Cambodia’s decades of political conflict under the reign of CPK may have resulted in lower rates of deforestation due to the lack of timber demand prior to the 1980s (Boutros-Ghali, 1994; Le Billon, 2000). Studies on the political ecology coupled with this study can help shed light on the transition from war to peace or the nature of different industries and their impacts on land cover and land use change.

In Cambodia, we calculated a total of 742 km² for the period of 1972 to 1977. This is a more conservative estimate compared to other studies that inventoried a total of 946 km² for the same period without using remote sensing (Cambodia Land Cover Atlas 1985/87–1992/93, 1994; Cambodia Forestry Policy Assessment, 1996; Ministry of Environment, Kingdom of Cambodia, 2009). Following this time period, one study estimated a total of 758 km² for 1989, while another study mapped a total of 646 km² by 1996 (Kozhikkodan Veettil and Quang, 2019; Bunting et al., 2022). Over the course of 1996 to 2016, different studies estimated a mangrove loss of 208–300 km². When comparing the new 1970s baseline to the estimate by Bunting et al. (2022) in 2020, this study indicates an additional loss of 115 km² ± 174 km². Many of the drivers of degradation and change have been attributed to shrimp pond aquaculture; salt pan and charcoal production; and infrastructure development (Kozhikkodan Veettil and Quang, 2019), but we suggest that the drivers of change were also very much political especially during the temporal period of this study (Boutros-Ghali, 1994; Le Billon, 2000).

It is important to note that several studies have indicated a high level of uncertainty on Cambodian mangrove forest estimates (Rizvi and Singer, 2016; Nop et al., 2017; Kozhikkodan Veettil and Quang, 2019). Although there is some uncertainty, these statistics show a trend of little change between the 1970s era up to 1996 and may serve as additional support to other claims on the political drivers of mangrove persistence. During the period of 1975 to 1979, the CPK (Boutros-Ghali, 1994; Le Billon, 2000) was the ruling regime in Cambodia. During this time, the country experienced severe famine, deaths associated to the lack of medicine, the proliferation of disease, and mass genocide which led to the deaths of up to 1.5 to 2 million people (Locard, 2005). These extremely difficult circumstances indirectly enabled forest stability and even gains (Le Billon, 2000) throughout the 1970s, which helped almost two-thirds of the country to be completely forested by the early 1980s (Cambodia Land Cover Atlas 1985/87–1992/93, 1994). Also, the aquaculture industry was not actively introduced until 1960 in Cambodia and was relatively inactive during the time of the CPK conflict. The active political conflict coupled with the lack of aquaculture activities are the likely drivers of mangrove stability seen between the 1970s to the 1990s.

Our estimate of mangrove extent from the 1970s for Thailand (5,407 km²) were somewhat higher than estimates from previous studies that mapped an extent with a range of 3,127–3,679 km² (Aksornkoae, 2012; Pumijumnong, 2014; Charupphat and Charupphat, 1997; Klankamsorn and Charuppat, 1982; Barbier and Cox, 2002; Naito and Traesupap, 2014). However, the large losses in mangrove extent that we observed between the 1970s and 1996 agree with other studies (Charupphat and Charupphat, 1997; Klankamsorn and Charuppat, 1982; Aksornkoae, 2012; Naito and Traesupap, 2014). One such study that documented the economic and demographic drivers of mangrove losses over the course of 1975 to 1996 in Thailand indicated that almost half (46%) of all mangroves were lost to coastal shrimp farming (50–65%) and the increased demand for land in coastal areas due to urbanization and agriculture during the period of 1979 to 1996 (Barbier and Cox, 2002; Spalding et al., 2010). This study indicates that the losses were 12% greater (58%) over the same time.

Mangrove losses associated with Thailand were primarily related to the demand for land development, aquaculture, economic incentives, and policy failures, but only up until industries started to aggressively expand. The studies by Barbier and Cox (2002) and Bantoon (1994) observed how even though aquaculture was introduced as early as 1974, the industry began to have severe mangrove loss impacts after 1985 due to Japanese demand (Bantoon, 1994; Barbier and Cox, 2002). In addition, sustained economic growth caused coastal populations to grow, boosting the demand for urbanization and economic crowding around mangrove areas. This caused shrimp production to go from 15,000 metric tonnes (KMT) in 1985 to 264,000 KMT in 1994 (Pednekar, 1998; Barbier and Cox, 2002). It also caused the shrimp farm area to rapidly expand between 1983 and 1996 and the number of farms to increase from 3,779 to 21,917. This boom in the aquaculture industry in addition to increased urbanization are the likely contributors to the severe mangrove losses seen throughout Thailand.

Previous estimates of mangrove extent and change in Myanmar exhibit a large amount of variation, highlighting the need for better data for this region (Aung, 2007; Spalding et al., 2010; Thant et al., 2012; Webb et al., 2014; Giardino et al., 2016; Veettil et al., 2018; Alban et al., 2020). Some of these claims documented that extreme overexploitation began as early as the Second World War to satisfy the demands of the military with the worst forest overexploitation occurring over the period of 1949 to 1972. According to Oo (2002) and Kyi (1992), mangrove forests in the Ayeyarwady delta decreased from 2,345 km² in 1954 to 1,786 km² in 1984. This study speculated losses that are much higher than our findings which established an area of 9,272 km² in the 1970s. More recently, Bunting et al. (2022) mapped a total of 5,821 km² of mangroves across the entire country in 1996, while Alban et al. (2020) mapped a total of 13,233 km² for the same year. These contradictory statistics demonstrate the need for region-specific mapping approaches in lieu of sub-setting a global dataset to report on country specific mangrove extent (Estoque et al., 2018; Alban et al., 2020). Also, Alban et al. (2020) estimated higher rates of mangrove deforestation in Myanmar, with 60% of mangroves permanently or temporarily lost between 1996 and 2016 due to the cultivation of rice, oil palm, and rubber in addition to urbanization.

Other studies that were done during the 1990s indicated additional land use drivers of change were to blame for the severe losses that were also found in Myanmar in the three main mangrove regions: Ayeyarwady, Tanintharyi, and Rakhine. According to Oo (2002), the main objective of mangrove management was fuel-wood and charcoal production. Then during Myanmar’s period of insurgency (1949–1972), the forest department was not able to effectively use forest management at large. However, mangrove species were not as affected by commercial demand, but by local demand. For example, the more heavily populated area of Yangon relied heavily on mangrove for charcoal and firewood production from the Ayeyarwady area, resulting in heavy losses. The Rakhine state on the other hand is an area with lower population and was speculated to have minimized losses according to Oo (2002) in the early 2000s. Higher populations and the demand for charcoal and fuel wood production therefore led to more extensive mangrove losses according to these studies. Following such losses, the Forest Department started intensive mangrove planting projects in 1975 followed by strict prohibitions of mangrove-derived charcoal and firewood after 1993. An important distinction that the study by Oo (2002) made is that many of the areas were experiencing losses due to the lack of electricity and need for energy which was an indirect driver of the fuel wood and charcoal production industry before prohibition.

Much of the losses of mangroves in our study region were driven by economic concerns. The Ayeyarwady delta of Myanmar experienced losses to mangrove charcoal and firewood production, but this resource was used to address the complete lack of electricity in the adjacent area of Yangon. This study indicated that almost half of mangrove area was lost by 1996, but we also urge researchers to have a sensitive perspective on the drivers behind the loss. Yes, these losses were extensive, but these results should be considered within the context of the people of Myanmar who had an urgent need to address their energy security, or the coastal communities of Thailand who pursued the economic benefits of the shrimp industry or urbanization, and the people of Cambodia who worked diligently towards reconstruction after the Khmer rouge. This study allowed us to not only establish a new baseline that would better inform current understandings of mangrove change before the 1990s, but it aided our understanding of the different needs that the people and their governments were trying to meet. We hope that this new baseline and conversation on the political ecology can serve as an example of good research practices in trying to understand mangrove change dynamics without forgetting about the human element.