Assessment of land cover changes using sentinel-2 satellite image data: A case study of Thanh Hoa coastal area, Viet Nam

Land use and management decisions heavily rely on the evaluating changes in land cover. Remote sensing techniques are becoming increasingly reliable and applicable, making them particularly useful for assessing and monitoring changes in land cover. In this study, the authors used Sentinel-2 MSI optical satellite image data and machine learning algorithms to classify and evaluate land cover changes in Thanh Hoa province’s coastal area between 2015 and 2023. Our research findings showed that Sentinel-2 MSI satellite image data can accurately interpret and classify land cover, with Kappa values ranging from 0.892 to 0.907. Furthermore, our findings indicated an increase in the area covered by build-up class. Meanwhile, vegetation cover and water surface class tend to decrease, especially the sharp decline of surface water. Research results help local policymakers develop land use plans in the direction of sustainable economic and social development.


Introduction
Researching land cover changes is important for making recommendations on rational and sustainable land use (Ademola et al., 2010).Land cover change is an extremely important factor in observing and studying greenhouse gas emissions and climate change.However, assessing and monitoring land cover changes over a large area cannot be done using traditional methods.It needs the support of remote monitoring data, which is remote sensing images from satellites.The global land cover product from satellite images that are considered to have the highest accuracy today is Landsat and Sentinel-2 satellite image data.In particular, using these free data sources can also save time, costs and efforts.
In Vietnam, applying GIS and remote sensing technology to evaluate and monitor land cover changes has also been deployed quite a lot in recent years.Remote sensing and GIS data results have been used in the interpretation and classification of land cover with high accuracy with Kappa coefficients from 0.71 to 0.89 (Pham Huu Ty et al., 2021).Modeling results in this technology reach an accuracy rate of nearly 70% (Dao Van Khanh and Nguyen Trong Truong Son, 2019).Remote sensing data with multitime characteristics, fast processing and wide area coverage is an effective tool for monitoring land use changes accurately and quickly.The use of remote sensing and GIS images also allows for editing and supplementing necessary data that cannot be conducted during investigation, survey, and measurement activities (Tran Thi Ly et al., 2019).However, these studies also have pointed out some difficulties when 1345 (2024) 012026 IOP Publishing doi:10.1088/1755-1315/1345/1/012026 2 using Landsat satellite images to evaluate the quality of image interpretation in different regions.The reason for this is that each locality records images with different metrics and image quality.
Most of the above studies used Landsat satellite images with low image resolution (30m), leading to significant/noticeable errors in interpreting land use/land cover.Therefore, the study proposes a method to apply satellite images with higher resolution than Landsat images to evaluate an area, which will have great practical significance.Some types of optical satellite images that users can obtain for free are Landsat, ASTER, MODIS, and, most recently, Sentinel-2 satellite data.Although Sentinel-2 is a new imager that was launched into orbit in 2015, it has attracted a great deal of attention from the scientific community as well as practical applications because of its multi-spectral band characteristics and relatively high spatial resolution (10m).
LULC classification using optical satellite imagery has recently been successfully performed using a machine learning approach, which improves accuracy compared to more conventional classification techniques (Talukdar et al., 2020).Generally, there are four types of machine learning algorithms: supervised, unsupervised, semi-supervised, and reinforcement learning (Xie et al., 2022), nonetheless, supervised and unsupervised learning are the most popular.Supervised learning techniques include support vector machines (Cortes et al., 1995), random forests (Breiman, 2001), classification and regression trees (Breiman et al., 1984), artificial neural networks (Tim Hill et al., 1994), while other techniques unsupervised learning includes fuzzy c-means algorithms (Saman et al., 2014), K-means algorithms (Abbas et al., 2016).
From the analysis of the above studies, remote sensing technology is a suitable and effective tool for monitoring and supervising land cover/land cover changes in the coastal area of Thanh Hoa province in the period 2015 -2023.The period 2015 -2023 provides favourable conditions because Sentinel-2 satellite images are only collected in a shorter period than Landsat images.On the other hand, due to local practice, land use status maps are prepared every 5 years in association with land inventory; thus, the period 2015 -2020 is completely suitable to evaluate changes, and the period 2020 -2023 is to prepare for the following land use planning period.The research focuses on solving two objectives: (1) Using Sentinel-2 MSI satellite images to classify land cover in the coastal area of Thanh Hoa province with high accuracy in 3 years 2015, 2020, and 2023; (2) Assessing the change in land cover in the coastal area of Thanh Hoa province in the period 2015 -2023.

Materials
The study uses remote sensing data such as Sentinel-2 satellite images to help improve image resolution.This has been proven in previous research, showing that the classification accuracy from the NDBaI index using Sentinel-2 satellite images increased by about 6% compared to Landsat 8 satellite images (Trinh Le Hung, 2020).In another study, it was also shown that Sentinel-2 images have higher accuracy, with an overall accuracy of 86.7% and a Kappa coefficient of 0.73, compared to Landsat-8 images, with an overall accuracy of 73.5% and Kappa coefficient of 0.47.Therefore, Sentinel-2 satellite images have greater applicability in creating land use status maps (Vo Quoc Tuan et al., 2021).Among the satellite images used to classify land cover, Sentinel-2 images are used more due to their free availability, short acquisition cycle, multiple spectral bands, and relatively high spatial resolution (Duan et al., 2019).From there, it shows that studying land surface cover changes in the coastal area of Thanh Hoa province in the period 2015 -2023 using Sentinel-2 MSI satellite images is completely reasonable and accurate to help improve efficiency in classifying land surface cover, contributing to the correct assessment of their changing trends.
Sentinel-2A was launched into orbit on June 23, 2015.This is a satellite equipped with a multispectral image acquisition with 13 spectral bands (443 nm-2190 nm), swath width 290 km, spatial resolutions 10 m (4 visible and near-infrared bands), 20 m (6 red-edges/shortwave-infrared) and 60 m (3 atmospheric correction bands).When the second satellite (Sentinel-2B) is put into use, both will have a repeat cycle of 5 days.This data shows higher spatial resolution than Landsat 8 satellite images.

3
The data used in the study is Sentinel-2 satellite image data with 10m spatial resolution, 5-day time resolution with 12 spectral bands.Images are downloaded on the Google Earth Engine cloud computing platform.To evaluate the change in land cover in Thanh Hoa province, the authors collected 03 Sentinel-2 satellite images in 2015, 2020, and 2023 within the boundaries of 3 typical coastal districts in Thanh Hoa province including Hoang Hoa district, Sam Son city, and Quang Xuong district.The study selected images with the best quality, with little influence of clouds.All processes of selecting and filtering images by time, location, and cloud cover to select images with the best quality suitable for the research period are performed on GEE software (Google Earth Engine is a free platform for cloud-based geospatial analysis).Information on Sentinel 2 images was collected for the presented study (Table 1).

Map construction method.
With Sentinel-2 satellite image data of the coastal area of Thanh Hoa province (Hoang Hoa district, Sam Son city and Quang Xuong district), the study selected sample data for 5 classes of LULC objects.After classification, the authors built a map using ArcGIS 10.8 software.

Methods of evaluating change.
Research on using image subtraction algorithm in assessing land surface cover changes in the coastal area of Thanh Hoa province.A change assessment model based on 3 map layers has been built.In this method, the model will browse each LULC type and determine the number of converted pixels of each LULC type at each time point to calculate the number of pixels changed and the percentage change for each type of LULC.

Establishment of land cover map of coastal area of Thanh Hoa province
Land cover map data of coastal districts of Thanh Hoa province in 2015, 2020, and 2023 was converted to spatial data and built a database after grouping into 5 types of objects: water bodies forests, grasslands, croplands, and built-up.Based on the results obtained in Table 2 of the above 5 object classes, croplands account for the majority of the area (>50%) in the coastal area of Thanh Hoa province with 340.12 ha (accounting for 68.44%) -in 2015; 302.69 ha (accounting for 60.96%) -in 2020 and 295.11 ha (accounting for 59.38%) -in 2023.This completely explains the suitability of land use with the topographic conditions of the coastal area of Thanh Hoa province.The coastal districts of Thanh Hoa have many rivers flowing through them, with a long coastline and many creek mouths, creating a river-sea alluvial plain.Alluvial, sandy, and saline soil account for a large area of the soil types here.These types of soil are very favourable for agricultural development, especially growing rice, and short-term crops.
Thanh Hoa coastal region has much potential for developing a diverse economy; however, agricultural production still accounts for a high proportion of more than 60% of the region's GDP structure.This economic sector depends greatly on nature, especially land resources.
The obtained Sentinel-2 images show that cropland cover is largely concentrated in Hoang Hoa and Quang Xuong districts, mainly used for growing tubers and fruits.This is the largest crop-growing area in Thanh Hoa province.Of the total area of sandy sealand, land capable of cultivating brackish water aquaculture is 10,386 ha.Among the three research areas (Hoang Hoa, Sam Son, and Quang Xuong), Hoang Hoa has the highest natural area (20,220 ha) and is also the district with the highest croplands area in all three regions (about 12,968.4ha); the lowest is Sam Son city with a natural area of 1,784 ha (with 554.31 ha of croplands).However, by 2015, the natural land area of Hoang Hoa and Quang Xuong had decreased compared to previous years (due to the planning to expand the area of Thanh Hoa city (6 communes of Hoang Hoa district and 5 communes of Quang Xuong district merged into the city), agricultural production is still the main economic activity in Thanh Hoa coastal districts (except Sam Son city).The data table also shows that the area covered by vegetation is mainly in the coastal area with an area of 383.66km 2 (accounting for 77.20% of the total area of the study area) -in 2015; 344.08km 2 (accounting for 69.29%) -in 2020 and 316.97km 2 (accounting for 63.77%) -in 2023.Next is the area covered by built-up, with an area of 89.09km 2 (accounting for 17.93% of the study area); 131.27km 2 (accounting for 26.41%) -in 2020 and 162.88km 2 (accounting for 32.77%) -in 2023.And finally, the area covered by water body with a sharp decreasing trend from 24.23 km 2 (accounting for 4.87%) -in 2015; down to 21.62 km 2 (accounting for 4.35%) -in 2020 and lowest in 2023 with 17.13 km 2 (accounting for 3.45% of the total land cover area in the study area).

Assessing land cover changes in the coastal area of Thanh Hoa province in the period 2015 -2023
Comparison results show that: in 5 years, the area of built-up tends to increase sharply, with 42.18 km2 in the period 2015 -2020, equivalent to an increase of 1.47 times (average area increased to 8,436 km2/year).This is completely consistent with the general development level of the Vietnamese economy.Built-up is unevenly distributed across administrative units and unevenly distributed between plains and mountainous areas.Built-up is mainly concentrated in cities, towns, and coastal towns, along rivers, and sparsely in mountainous areas.According to the 2020 Statistical Yearbook, Thanh Hoa has about 3.66 million people, the third largest population in the country, after two cities, Ho Chi Minh city and Hanoi city, ranked first in all provinces.The average population density of Thanh Hoa is 330 people/km 2 .Thanh Hoa city alone has a density of 2,400 people/km 2 , Hau Loc, Hoang Hoa, and Quang Xuong districts have a density of over 1,100 people/km 2 .Meanwhile, the density is low in mountainous districts such as Muong Lat, Quan Son, and Quan Hoa, from only 39 to 46 people/km 2 .The main causes of the above differential distribution of built-up include the uneven distribution of natural resources, the system of socio-economic facilities, the infrastructure system to serve life, and residence history.The natural population growth rate is estimated at 0.53%.That explains why analysis from Sentinel-2 MSI satellite images shows that the built-up covers a large and increasingly large area in the study area.
The comparison results also show that changes in cropland cover, vegetation cover, and water body all tend to decrease during this period.Of which, cropland cover decreased the most with 37. Vegetation cover has decreased sharply recently, especially in grasslands.The area of wind and sand protection forests is seriously decreasing, mainly due to human activities from converting forestry land use purposes to non-forestry land use purposes such as developing service tourism, urbanization, and aquaculture.At the same time, there is a lack of silvicultural techniques in protecting and developing forest funds.
Comparison results in the period 2020 -2023 show that in 03 years, the area of built-up cover still tends to increase strongly with 32.77 km 2 , equivalent to an increase of 1.24 times (an average increase of the area of 10.5 km 2 /year).This increase is even faster than the previous period (2015 -2020).This leads to increased risks in coastal areas because safe areas are increasingly densely populated, so new built-up areas must be concentrated in areas with twice the risk of flooding than other areas.Urban areas IOP Publishing doi:10.1088/1755-1315/1345/1/01202610 have existed for a long time.On the other hand, the concentration of construction land and urbanization in coastal areas increases wastewater and solid waste, affecting the marine environment.The vegetation and water in oceans, seas, lakes, rivers and glaciers are still decreasing due to the increase in built-up areas.The research is currently using the Support Vector Machine algorithm for the classification process.This is a machine learning algorithm that was proposed in 1995.Therefore, the object classification process is not highly accurate.In recent times, the development of neural networks has made great strides in image classification.Some models such as CNN, RNN, DNN have shown effective classification results.Research will continue to develop classification and evaluation models the future.

Conclusion
Using Sentinel-2 satellite image data for land cover interpretation and classification, the study has yielded the following conclusions: During the period from 2015 to 2023, the land cover in the coastal area of Thanh Hoa province is anticipated to undergo significant changes.Notably, the built-up area is expected to experience a rapid expansion.Conversely, water bodies, forests, and croplands are projected to decrease by 73.9 km2 over the entire 8-year cycle.This transformation can be attributed to the fact that the coastal region serves as a focal point for several of the country's key economic sectors, including industry, services, and tourism.However, the surge in coastal construction land also presents numerous challenges concerning the effective management and sustainable utilization of land, natural resources, and environmental protection in coastal areas.

Figure 4 .
Figure 4. Flow chart for assessing land cover changes from Sentinel-2 MSI images2.2.2.Object classification method.The authors use the object-oriented classification method to classify land cover.The object-oriented classification method is a modern and highly effective classification method when classifying land cover compared to other traditional classification methods because some characteristics of objects as well as the interaction of that object with other objects are calculated during the classification process.To segment objects and extract features for object-oriented classification, we use the SNIC segmentation algorithm combining the Gray-Level Co-occurrence Matrix (GLCM) and the principal component analysis (PCA) algorithm in the study.The study conducted sampling for the classification process including the following objects: water body, forest, grasslands, croplands, built-up.Samples are in polygon shape and are taken throughout the entire study area.Finally, the sample data set is randomly divided into a training data set and a test data set in a 7:3 ratio.After sampling, the study conducts classification.Based on the characteristics of the object classes, after segmentation, take training samples for land cover objects.The study conducts land cover classification based on the object-oriented classification method combined with the Support Vector Machine (SVM) algorithm.

7 Figure 5 .Figure 6 .
Figure 5. Image classifying land cover objects in the study area in 2015

Figure 7 .
Figure 7. Image classifying land cover objects in the study area in 2023

Figure 8 .
Image changes land cover objects during study periods: (8a) -Changes of land cover in the period 2015 -2020; (8a) -Changes of land cover in the period 2015 -2023.
15km 2 within 5 years (2015 -2020), equivalent to a decrease of 7.43km 2 /year.This result is consistent with Decision No. 326/QD-TTg dated March 9, 2022 of the Prime Minister on the allocation of national land use planning targets for the period 2021 -2030, with a vision to 2050, 5-year national land use plan 2021 -2025.

Table 1 .
Sentinel-2 image data information used in research

Table 2 .
Statistics of land cover area after grouping

Table 3 .
Comparison results of land cover area changes in coastal Thanh Hoa province in the period 2015 -2023