Thirty-Year Dynamics of LULC at the Dong Thap Muoi Area, Southern Vietnam, Using Google Earth Engine

: The main purpose of this paper is to assess the land use and land cover (LULC) changes for thirty years, from 1990–2020, in the Dong Thap Muoi, a ﬂooded land area of the Mekong River Delta of Vietnam using Google Earth Engine and random forest algorithm. The speciﬁc purposes are: (1) determine the main LULC classes and (2) compute and analyze the magnitude and rate of changes for these LULC classes. For the above purposes, 128 Landsat images, topographic maps, land use status maps, cadastral maps, and ancillary data were collected and utilized to derive the LULC maps using the random forest classiﬁcation algorithm. The overall accuracy of the LULC maps for 1990, 2000, 2010, and 2020 are 88.9, 83.5, 87.1, and 85.6%, respectively. The result showed that the unused land was dominant in 1990 with 28.9 % of the total area, but it was primarily converted to the paddy, a new dominant LULC class in 2020 (45.1%). The forest was reduced signiﬁcantly from 14.4% in 1990 to only 5.5% of the total area in 2020. Whereas at the same time, the built-up increased from 0.3% to 6.2% of the total area. This research may help the authorities design exploitation policies for the Dong Thap Muoi’s socio-economic development and develop a new, stable, and sustainable ecosystem, promoting the advantages of the region, early forming a diversiﬁed agricultural structure.


Introduction
River deltas are considered to be some of the most favorable areas for crowded residential communities in the world, as they cover only around 5% of the total land area globally but are home to more than 500 million people [1]. However, these deltas have been recognized as the main places for significant land use and land cover (LULC) changes due to population growth, urbanization, anthropogenic activities, and socioeconomic development [2]. Over the last decade, these changes have affected various global issues, both positive and negative, including terrestrial ecosystems [3], land surface temperatures [4], regional climates [5], soil quality [6], hydrological regimes [7], flooding [8,9], other natural hazards [10,11], and especially food security [12,13]; therefore, studying LULC changes and assessing driving factors to balance the correlation between the human needs and negative consequences are vital factors for sustainable development [14]. The Dong Thap Muoi climate is characterized by tropical monsoons, being hot, sunny, and high in humidity. This area has only two distinct seasons: rainy and dry seasons. The total yearly precipitation fluctuates in the range of 1300-1700 mm, with the rainy season from May to October accounting for up to 90% of the yearly precipitation [39]. In general, the whole region has a high annual average temperature and is stable in space and time. According to the climate statistics from 1978-2019 at the Cao Lanh station in the study are, the highest and lowest mean annual temperatures were 26.7 and 27.9 °C, respectively. April and January are the months with the highest and lowest average temperatures, respectively.
According to the statistical analysis at Cao Lanh station during 1978-2019, the lowest measured temperature was 15.8 °C in 1986, whereas the highest temperature observed was 37.4 °C in 1983. Regarding the wind regime, the southwest wind from the Gulf of Thailand, usually from May to November, is dominant. This wind is steamy, and therefore causes rain. Meanwhile, the prevailing northeast monsoon season from November to January is usually dry [36]. The average wind speed is 2-2.5 m/s. The terrain of Dong Thap Muoi is generally flat, with around 73% of the total area between altitudes of 0.5 and 1.25m, whereas areas with an altitude over 2 m occupy only about 7% and the lowest areas with altitudes of 0-0.5m account for only 0.57% of the total area [36]. Regarding the soil, acid sulfate soils cover approximately 39% of the total area, followed by alluvial soil (34.71%), gray soils (16.10%), sandy soils (0.51%), and peat soils (0.02%) [36]. Due to its topographical characteristics, Dong Thap Muoi is a closed flood plain and is strongly influenced by the semi-diurnal tide regime of the East Vietnam Sea. Thus, regular floods, droughts, and acid sulfate problems are major environmental issues, affecting agricultural production and people's lives in the area. Although Dong Thap Muoi covers only 17.7% of the Mekong River Delta area, it occupied 36.7% of the delta's total flooded area [36]. In a normal year, drought and saline intrusion occur between January and April, followed by acid sulfate problems, which occur from May to July, while August to November is the flood season [37,38].
The Dong Thap Muoi climate is characterized by tropical monsoons, being hot, sunny, and high in humidity. This area has only two distinct seasons: rainy and dry seasons. The total yearly precipitation fluctuates in the range of 1300-1700 mm, with the rainy season from May to October accounting for up to 90% of the yearly precipitation [39]. In general, the whole region has a high annual average temperature and is stable in space and time. According to the climate statistics from 1978-2019 at the Cao Lanh station in the study are, the highest and lowest mean annual temperatures were 26.7 and 27.9 • C, respectively. April and January are the months with the highest and lowest average temperatures, respectively.
According to the statistical analysis at Cao Lanh station during 1978-2019, the lowest measured temperature was 15.8 • C in 1986, whereas the highest temperature observed was 37.4 • C in 1983. Regarding the wind regime, the southwest wind from the Gulf of Thailand, usually from May to November, is dominant. This wind is steamy, and therefore causes rain. Meanwhile, the prevailing northeast monsoon season from November to January is usually dry [36]. The average wind speed is 2-2.5 m/s.

Remote Sensing Images, Ancillary Maps, and Fieldwork Data
In order to cover the Dong Thap Muoi area, a total of 128 scenes with 30 m resolution in the 04 paths and rows (125/052, 125/053, 126/052, and 126/053) were derived from the freely available data catalog in Google Earth Engine (GEE) (https://developers.google. com/earth-engine/datasets/catalog (Accessed date: 20 September 2020)). Among the 128 scenes, 98 Landsat-5 scenes were taken from 1990, 2000, 2001, and 2010, whereas 3 Landsat 8 OLI (Operational Land Imager) scenes were taken from 2020. Herein, Landsat 7 images were not considered due to parallel stripe problems. We selected the images from the dry seasons (January to April) of the aforementioned years to minimize the influence of the cloud coverage and the effects of moisture and evaporation on the ground. Nevertheless, the area was still influenced by cloud cover for the years 2000 and 2010, so in order to fix this we used 23 and 14 Landsat-5 scenes for the years 2001 and 2011 for the pixel value recovery processing to derive the full images from 2000 and 2010 for the Dong Thap Muoi area. A detailed description of the Landsat images used in this project is shown in Table 1. river networks, residential areas, and land uses, were also used. In this project, the ALOS DEM with 30 m resolution was downloaded from the Japan Aerospace Exploration Agency (JAXA) website (https://www.eorc.jaxa.jp, assessed on 2 April 2020). The boundaries, road networks, and river networks were derived from the national spatial data infrastructure of the Vietnam Ministry of Natural Resources and Environment (V-MONRE) [40]. For 1990, 2000, 2010, and 2020, 1:2000 cadastre maps provided by the local authorities, land use status maps, and topographic maps compiled by V-MONRE were used for comparisons with fieldwork results. The other ancillary data used in this project are shown in Table 2. The fieldwork was conducted in May and June 2020 using a Trimble handheld GNSS receiver, while an iGeoTrans app [41] installed on an iPad was used to collect samples using the random sampling method [42]. A sampling unit of 30 × 30 m was adopted to the spatial resolution of the Landsat images. Herein, the study area's road network overlaid on Google Earth was used to support the designed field trips. A total of 6 individuals separated into 3 groups conducted these field trips, visiting and checking 422 specific sites covering all of the LULC classes in the Dong Thap Muoi area. These specific sites are spread over 700 km of the roads in the study area. The distribution of the fieldwork sites is shown in Figure 2, while some image samples from the Landsat images are shown in Figure 3. Photos of some LULC classes taken during the fieldwork in the study area are shown in Figure 4. During the field trips, interviews were also carried out with local leaders of communes, village heads, forest ranger staff, and elders.

Image Pre-Processing
As mentioned above, the Landsat images for the aforementioned years were derived from the freely available GEE. These images were processed and corrected to surface reflectance values, which included corrections for the solar zenith and irradiance, sun, and atmosphere [43,44]. Because the Dong Thap Muoi is a flat area with very low terrain variation, topographic correction with a DEM for this area was not necessary [15,45]. In this project, the pre-processing was carried out in 8 steps; (1) creation of image data file;

Image Pre-Processing
As mentioned above, the Landsat images for the aforementioned years were derived from the freely available GEE. These images were processed and corrected to surface reflectance values, which included corrections for the solar zenith and irradiance, sun, and atmosphere [43,44]. Because the Dong Thap Muoi is a flat area with very low terrain variation, topographic correction with a DEM for this area was not necessary [15,45]. In this project, the pre-processing was carried out in 8 steps; (1) creation of image data file; (2) cloud filtering; (3) creation of average image of study area; (4) selection of image bands and computation of NDVI (Normalized Difference Vegetation Index) and NDWI (Normalized Difference Water Index) for classification; (5) preparation of LULC classes, followed by training and validation of samples; (6) classification LULC using the random forest approach; (7) accuracy assessment; (8) post-processing and analysis.
For the classification, five bands (blue, red, green, NIR (Near Infrared band), and SWIR (Short-Wave infrared)) of Landsat-5 were used, whereas for Landsat 8 OLI, NDVI and NDWI were additionally computed and utilized. For the post-processing, the final classified images were geometrically corrected to the VN-2000 (Viet Nam 2000) system (zone 48, UTM map projection, ellipsoid WGS84) [46] using the nearest-neighbor sampling algorithm [47]. The purpose of this geometrical correction was to ensure the best alignment between these images and the other spatial data. It is noted that VN-2000 is the national coordinate reference system of Vietnam, which has been officially used since the year 2000 [46]. Herein, topographic maps at 1:50,000 scale (refer to Table 2) were used for this image-to-map registration, and root mean squared error (RMSE) accuracy values below 0.5 pixels [15] were accepted

Determination of the LULC Classes, Training and Validation of Samples
In this research, the LULC classes for the Dong Thap Muoi area were determined based on the regulations for mapping and updating land use status maps in Vietnam [48], on our fieldwork, and on the land use statistics available for this area. In addition, the instructions for land cover classification of the Food and Agriculture Organization (FAO) [49] and The United States Geological Survey (USGS) land cover database [50] were considered. As a result, a total of 8 broad LULC classes (paddy, built-up, water, forest, orchards, crops, unused land (bare land, wetland, shrubs), and others) were determined (Figures 3 and 4  and Table 3). We used only 8 LULC classes here to raise the probability of the changes are true, whereas ambiguities of categorical differentiation could be reduced. For supervised classification, it is necessary to prepare training and validation samples. In this project, we selected these samples for the years 1990, 2000, 2010, and 2020 from 1:50,000 topographic maps, 1:50,000 land use status maps, and 1:2000 cadastre maps of the study area. Herein, we used the procedure proposed by Fortier et al. [51] with the random stratification method [52] for the selection. First, Landsat images for 1990, 2000, 2010, and 2020 were placed over these maps in ArcGIS Pro. Then, invariant areas in the eight LULC classes were determined and polygons were drawn in the centers of these areas. Finally, the labels for LULC classes were checked using the three map types listed above. As a result, a total of 1397 samples in the eight LULC classes for the years 1990, 2000, and 2010 were generated using the three ancillary maps for training models with the random forest algorithm. For 2020, only 975 samples were generated using the above ancillary maps, whereas 422 samples were derived through our field trips in June 2020.

Landsat Image Classification and Accuracy Assessment
Due to the 30 m resolution of the Landsat images used, a pixel-based supervised classification approach was adopted to derive LULC classes for the Dong Thap Muoi area. Since LULC class quality depends on the method or algorithm used and decision tree ensembles have proven to be efficient due to their high accuracy and resistance to noise [53][54][55], in this research the random forest approach was selected.
The random forest method is a powerful classification scheme proposed by Breiman [56], which has been successfully used in various LULC classification projects [53,[57][58][59][60][61]. As pointed out by Sheykhmousa et al. [57], the main advantage of RF is that it can provide very high classification results, and in addition this algorithm is robust to outliers and works well in noisy environments [53]. The working mechanism for RF can be summed up as follows: first, using the LULC training samples, this algorithm separates the training samples into subsets with the use of the bagging technique [62]. Then, each subset is utilized to construct a subtree classifier. Finally, all subtree classifiers are aggregated to create the final forest model. Overall, the random forest model's classification power is strongly influenced by the number of subtrees used, as a result of its diverse characteristics; therefore, this parameter should be carefully considered. As suggested in previous works [63][64][65], a total of 500 subtrees were used in this project.
For the accuracy assessment of the LULC maps in this project, both qualitative and quantitative methods were adopted. For the qualitative assessment, the LULC maps for the years 1990, 2000, 2010, and 2020 and their corresponding Landsat images were overlaid over topographic maps, land use status maps, and cadastre maps, then visual checks and assessments were carried out. For the quantitative verification, the LULC map quality was checked using various well-known and widely accepted statistical measures (i.e., kappa index; user, producer, and overall accuracy values), as mentioned in [42,66]. A detailed explanation of these statistical measures can be found in [42,47].

LULC Estimation
The final trained random forest models were used to derive the spatial patterns for eight LULC classes (paddy, built-up, water, forest, orchards, crops, unused land, and others) for the years 1990, 2000, 2010, and 2020. The results are shown in Figure 5 and Table 5.
It can be seen that the most dramatic change in the area is for the paddy classification, as it occupied only 935.0 km 2 in 1990 but steadily increased to 1874.5 km 2 in 2000, 2941.4 km 2 in 2010, and 3292.2 km 2 in 2020. Inversely, the unused land areas, located mainly in the north, crossing the border between Vietnam and Cambodia, covered 2111.9 km 2 in 1990 but were strongly reduced by nearly nine times by 2020 (236.0 km 2 ).
Showing a similar trend but at a slightly reduced level, the forest areas, which were mainly distributed in the northeast part in 1990 (1049.5 km 2 ), were reduced to 402.7 km 2 by the year 2020 and were located in fragmented clusters.

LULC Evolution Analysis
The evolution of the LULC classes in the Dong Thap Muoi area is shown in Table 6 and Figures 6-8. Among the eight classes, the paddy class showed the largest change, followed by the unused land class. During the period 1990-2020, the paddy area increased by 2357.2 km 2 , whereas the unused land area decreased by 1875.9 km 2 . These changes accounted for 32.3% (paddy) and 25.7% (unused land) of the total studied area. The highest annual rate of change for the paddy class was around 106.7 km 2 /year for the period 2000-2010, while for the unused land class it was 126.2 km 2 /year for the period 1990-2000 (Table 6).
These magnitude changes were followed by the forest (−646.8 km 2 ), built-up (432.0 km 2 ), crops (−292.4 km 2 ), water (87.7 km 2 ), and orchard (−64.3 km 2 ) classes between 1990 and 2020. These changes corresponded to 8.9, 5.9, 4.0, 1.2, and 0.9% of the total studied area. The annual rate of change for the forest was highest (−31.6 km 2 /year) in the period 1990-2000, which then reduced from 2000 to 2010 (−3.7 km 2 /year), but increased to -29.4 km 2 /year for the 2010-2020 period. Regarding the built-up class, the annual rate of change was high (20.8 km 2 /year) for the 2000-2020 period, especially from 2010 to 2020, where the rate reached 28.8 km 2 /year. The magnitude and annual rate changes for the other LULC classes are shown in Table 6

LULC Evolution Analysis
The evolution of the LULC classes in the Dong Thap Muoi area is shown in Table 6 and Figures 6-8. Among the eight classes, the paddy class showed the largest change, followed by the unused land class. During the period 1990-2020, the paddy area increased by 2357.2 km 2 , whereas the unused land area decreased by 1875.9 km 2 . These changes accounted for 32.3% (paddy) and 25.7% (unused land) of the total studied area. The highest annual rate of change for the paddy class was around 106.7 km 2 /year for the period 2000-2010, while for the unused land class it was 126.2 km 2 /year for the period 1990-2000 (Table 6).              Table 6 and Figures 7 and 8.
Detailed conversions of the LULC changes in the Dong Thap Muoi area are shown in Table 7, which presents the change matrix for the 1990-2020 period over six different time spans. Herein, the diagonal line denotes the number of the LULC classes showing no change, whereas the values above the diagonal line explain the expansion areas. The values below the diagonal line refer to the reduction areas.  Table 7, which presents the change matrix for the 1990-2020 period over six different time spans. Herein, the diagonal line denotes the number of the LULC classes showing no change, whereas the values above the diagonal line explain the expansion areas. The values below the diagonal line refer to the reduction areas.

Accuracy of the LULC Classification
The Dong Thap Muoi area is influenced by the tropical monsoon climate. Therefore, the collected images were affected by clouds and their shadows, although the Landsat images were collected in the dry season only. In this project, cloud filtering with the C Function of Mask (CFMask) algorithm available at GEE was used for pre-processing, however pixel reconstructions were still required to derive the "fresh" images. The overall accuracy of the LULC maps for 1990, 2000, 2010, and 2020 was higher than 83%, which satisfied the suggested thresholds mentioned by Were et al. [67] for LULC mapping with remotely sensed images. Among the LULC classes, the built-up class had the lowest accuracy, especially in the LULC maps for 2000 and 2010. Herein, built-up pixels were misclassified into the unused land class. However, this is a reasonable result because the built-up class was too fragmented before 2010. Thus, due to the topographical conditions, residence habits, and economic activity relating to wet rice cultivation, the inhabitants are mainly concentrated in towns and clusters along roads and irrigation canals. The misclassification was also due to spectral confusion between the crop and paddy classes (i.e., in 2000). Nevertheless, all LULC classes had user and producer accuracy rates higher than 70%, which are acceptable accuracy levels for the complex nature of LULC changes in the Dong Thap Muoi area. The kappa index values ranged from 0.835 to 0.889, indicating that the LULC models classified these images better than random images, which ranged from 83.5% to 88.9%.

LULC Changes
In the 1980s, irrigation canals were constructed to bring fresh water into the area, the process of exploitation was started, and the population began to increase. Due to the topographical conditions, residence habits, and the economic activity related to wet rice cultivation, the inhabitants are mainly concentrated in the towns and clusters along roads and irrigation canals. This analysis indicates that that the paddy, unused land, forest, and built-up classes showed the most important LULC changes in the Dong Thap Muoi area during the period 1990-2020. Herein, the unused land and paddy areas had the highest magnitude of changes. During the 30-year period (1990-2020), the unused land decreased by 1877.5 km 2 , and this change accounted for 25.6% of the total studied area. The highest annual rate of change was for the period 1990-2000 at around 126.3 km 2 /year, whereas the change was more stable for the period 2000-2020.
In 1990, the built-up area totaled for 18.8 km 2 , which accounted for 0.26% of the total Dong Thap Muoi area, however in 2020, the built-up area increased 24 times to 450.7 km 2 ( Table 6). The annual change rate was 14.4 km 2 /year. The highest change rate per year was for the 2010-2020 period (28.9 km 2 /year), followed by the 2000-2020 period (20.8 km 2 /year), the 2000-2010 period (12.7 km 2 /year), the 1990-2010 period (7.2 km 2 /year), and the 1990-2000 period (1.7 km 2 /year). Thus, the population increase was the main cause. The total population in 2001 was 2,463,377 people and the average population density was 337 people/km 2 [36], however in 2019 the total population was 2,594,211 people. From 2015 to 2019, the total population in the Dong Thap Muoi area increased by 18,730 people [68].
The population growth and increase in exploitation in the Dong Thap Muoi area were in line with the forest cover losses throughout the study period. The forest area totaled 1049.5 km 2 in 1990 and was consistently reduced to 402.7 km 2 in 2020 (Table 5).
In 1990, the water body area totaled 403.6 km 2 , which decreased to 292.6 km 2 in 2000 and then increased to 416.1 km 2 in 2010 and 491.3 km 2 in 2020 (Table 7). This is because in the 1990-2000 period, the expansion of agriculture practices was largely spontaneous and uncontrolled. Many wetland areas and natural rivers were used for these agricultural practices, which were considered a cause of flooding in the 2000s in the area, where the network of natural rivers and canals could not adapt to drain water. Then, flood control measures were developed [69,70] for the study area, including clearances of natural flow in rice cultivation areas, which increased the water bodies again.
The orchard land class increased from 1322.0 km 2 in 1990 to 1433.4 km 2 in 2000, and then reduced to 1186.0 km 2 in 2010 and increased again to 1257.7 km 2 in 2020. Overall, it was reduced by around 4.9%. For the crops classification, this LULC class sharply increased from 1462.4 km 2 in 1990 to 2085.8 km 2 in 2010 but was greatly reduced to 1170.0 km 2 in 2020. In total, it was reduced by around 20% during the 1990-2020 period. Herein, orchard and crop areas were lost mainly to the built-up and paddy areas. Thus, the policies of the local authorities regarding agriculture reconstruction [71,72] were the main reasons for the changes of these two LULC classes, aiming to achieve higher added value and sustainable development in this study area

Potential Driving Forces for the LULC Changes
During the years 1975-1980, most of the land in this area was unexploited and acid sulfate soil was dominant and difficult to cultivate [73]. In 1977 constructed started for the freshwater canal, namely the Hong Ngu-Vinh Hung canal, which has a length of 45 km and was completed in 1984. Exploitation then started to occur in this area, whereby people from other regions selected areas located on both sides of the canal to settle down, build houses, cultivate the land, and develop their new lives.
The LULC transformation and the development of Dong Thap Muoi entered a new phase in the years after 1986, which was strongly related to the socioeconomic renovation of the Vietnamese government, called the "Doi Moi" policy [74] in Vietnam. This was especially influenced by directive No. 515-TTg from the Vietnamese government enacted on 10 July 1997, which related to strengthening the implementation of exploitation activities and socioeconomic development in the Dong Thap Muoi area. As a result, the Dong Thap Muoi area was still mostly unexplored until nearly 1990, at which point unused land was dominant, however by 2020 the area had become almost fully exploited.
Changing from an area with harsh natural conditions and a sparsely distributed population, where most people made a living through self-sufficient production that was dependent on nature, Dong Thap Muoi has become an important rice production area. More importantly, irrigation, dike, and road networks have been expanded significantly [70,75,76], which favor the expansion of paddy cultivation. Additionally, the population has been reallocated better and people's cultural and social lives have been improved.

Concluding Remarks
This project assesses the thirty-year dynamics of land cover/land use changes in the Dong Thap Muoi area in Southern Vietnam for the period 1990-2020. A total of 128 scenes, topographic maps, land use status maps, cadastral maps, and ancillary data were adopted to derive the LULC maps. This is the first study to assess in detail the LULC changes in the Dong Thap Muoi area. From the results, we can draw the following conclusions: The paddy, built-up, water body, forest, orchard, crops, and unused land classes are the main LULC classes in the Dong Thap Muoi area; During the last 30 years, from 1990 to 2020, the paddy and built-up areas have increased greatly, whereas the forest and unused land areas have been significantly reduced. The paddy areas were transformed mainly from unused land areas; The LULC analysis results from this study may help the authorities design exploitation policies for the socioeconomic development of the Dong Thap Muoi area in the future; Google Earth Engine is a powerful tool for LULC classification; Future extensions of this research should focus on assessing the effects of the LULC changes in flood hazards and natural ecosystems in the Dong Thap Muoi area.