Spatial Dynamics Evolution of Land use for the Study of the Local Traditional Living Changes

DOI:

Human activities reflect land use in the way how people use the land for a specific purpose. Humans need the land for living, and they notice the house on the land. Therefore, land use data can be used to understand economic growth and behavior. Land use data provides information about how land is being used, such as whether it is being used for agriculture, industry, or residential purposes, and this can be used to analyze economic activity in those areas. For example, an increase in land use for industrial purposes may indicate growth in the manufacturing sector and suggest a growing local economy. Alternatively, a decrease in land use for agriculture may indicate a shift away from traditional farming practices and towards other economic activities, such as tourism or services [1]. Land use data can also be used to understand patterns of economic behavior, such as the relationship between land use and property values or the impact of land use on environmental factors like air and water quality. Understanding these relationships can inform policy decisions and guide investments in infrastructure and development. Overall, land use data can be a valuable tool for understanding economic growth and behavior [2] and [3], particularly when combined with other data sources and analysis methods. To be achieved this, the land activities that change over time have to investigate [4] [5] [6] and [7].

Land use and land cover (LULC) mapping derived from satellite imagery is a valuable source of information that offers current insights into various stages of transformation occurring on the Earth's surface [8]. In addition, the time series of imagery allows a seasoning of landscape and actual processes such as urban expansion, crop rotation, and deforestation [9]. Besides, the procedure of land’s surface can be detected or computed with satellite imagery through a multi-temporal perspective [10]. As a result, remote sensing has emerged as the most effective and reliable method for characterizing and analyzing land-use and land-cover conditions and changes, and its usage has increased significantly over time [11]. The Sentinel-2 satellites equipped with multispectral imaging instruments (MSI) offer high-resolution satellite data that is useful for monitoring LULC [12]. This missions are aimed at ensuring continuity and improving upon the accomplishments of the Landsat missions. [13].

Presently, the widely usage tool for satellite data accesses and processes is the Google Earth Engine (GEE). The GEE is a cloud-based computing platform that is easily accessible and capable of storing and processing large-scale geospatial data in the petabyte range. There is easily accessible and user-friendly with interactive data and processing algorithms [14]. Also, GEE provides high performance of cloud computation and multiple practical tools for carrying out the analysis of global geospatial big data. Currently, it is widely acknowledged as the most popular cloud computing platform in the fields of Earth and Environmental Science [15] that is commonly utilized for processing data related to various fields focused on environmental change, such as crop analysis [16], water resources [17], land cover mapping [18], disaster monitoring [19], climate change [20], soil properties [21], forest [22] and urbanization [23]. Due to the inability of traditional computing resources to handle multi-time series satellite images over large areas, the GEE service has emerged as an effective platform for supporting the analysis of global and regional land changes [24].

In this research, the LULC classification was conducted using the Sentinel-2 time-series imagery in conjunction with the random forest algorithm on the GEE platform. In addition, we conducted a change detection analysis to examine the LULC changes in Phetchaburi and Samut Songkhram provinces between the years 2019 and 2021. The cellular automata - Markov chain model (CA–Markov) deployed in the TerrSet software developed by Clark (formerly IDRISI), with the high reputation and widely acceptation. The software provides an excellent advantage and essential geospatial tools for the study of land-cover changes and probable prediction [25]. This study, itis used to predict the possible LULC in 2023. Additionally, the objectives of our study included the classification of LULC using a sustainable cloud-computing platform and training dataset, as well as the simulation of future land cover changes in our study area located in western Thailand through the utilization of the CA-Markov model.

2. Materials

2.1 Study Area

This research sites are located in Phetchaburi and Samut Songkhram provinces (Figure 1). These provinces are in the western region of Thailand (12°33′45′′N to13°31′7′′ N and 100°6′12′′E to 99°6′2′′ E) and cover the area of 6585.28 km². The natural border between Thailand and Myanmar is formed by the presence of high mountains and forests in the western region. Also, the east site is a coastline be a part of the gulf of Thailand. In addition, the Kaeng Krachan National Park, which was declared as a potential UNESCO Natural World Heritage site in 2013, is located in the western region [26].

2.2 Background of Study Site

The Phetchaburi province has a forestry of 57.24%, an agricultural area of 28.86%, and a build-up of 5.14%. The total population was approximately 480,000 people and the gross provincial product (GPP) was about 63,000 million bath and the GPP per capita was about 130,000 bath/year in 2016. The service was a largest economic section about 37,000 million baths and the industrial was the larger section about 18,000 million baths also the agriculture section was large economy about 7,700 million baths. The Samut Songkhram province has an agricultural area of 68.22%, a build-up of 15.24%, and forestry of 6.79% in 2020. The total population was approximately 194,000 people and the gross provincial product (GPP) was 20,400 million bath and GPP per capita was 105,000 bath/year. The service was a largest section about 12,600 million baths, and the industrial was larger section about 5,300 million baths, also the agriculture was large economic section approximately 2,400 million baths.

The main economic sectors in those area are the agriculture, fishery, and sea salt farm. The rapid expansion of industrial area in last decade is the most driving force to change the traditional economic sectors to industrial sectors which reflecting on land use change, as well as the local traditional living.

Figure 1: The study area of Phetchaburi and Samut Songkhram provinces

2.3 Dataset

The data used in this study are composed of two historic Sentinel-2 satellite imagery (2019 and 2021) with a spatial resolution of 10-meter. The imagery scenes that covered the study area irrespective of cloud cover were operated in the GEE platform. The image datasets were already geo-referenced, projected, and provided the atmospheric corrections. In addition, the land use datasets of Phetchaburi and Samut Songkhram province from the Land Development Department (LDD) were used to perform the training dataset for machine learning classification. These datasets were produced by visual interpretation from LULC experts of government. Also, the study area's ground truth dataset was observed randomly and covered in different LULC classes. The identified LULC classes were compared and matched with similar types observed in both the classified imagery and training dataset. Moreover, the independent variable datasets were collected to support the prediction model: the digital elevation model (DEM), the road network dataset, the stream dataset, and the land use dataset.

3. Methodology

In this section, we describe the major anticipated uses for the LULC changes in the future. The time-series raster dataset from Sentinel-2 imagery was classified using a machine learning algorithm on the Google Earth Engine platform. The primary workflow implemented in this study involved: (1) the mapping of 2019 and 2021 using the random forest classifier of satellite imagery, (2) an analysis of change detection, (3) predicting LULC changes using a CA-Markov Model in the year 2023 was a crucial aspect of the workflow in this study. The overall workflow illustrates in Figure 2.

3.1 Data Preparation

The time-series Sentinel-2 images were used in 2019 and 2021. Sentinel-2 (L2A) products were used in this study, which is corrected reflectance products [27] and [28], and are available on GEE. Cloud cover is a significant limiting factor in time-series datasets from satellite sensors, particularly in rainforest regions [29]. Hence, the library of sentinel-2 cloud probability product was utilized on the GEE platform to mask the cloud cover in all sentinel-2 images. We separate the period of imagery into three periods of each year: a beginning period (January-April), a middle period (May-August), and the end period (September-December). Each period consisted of five generated bands, namely; Blue, Green, Red, NIR, SWIR, and three indices named as NDVI, NDWI, and NDBI, that were resampled into 10-meter pixel sizes of all bands as demonstrates in Figure 3.

Figure 2: Overall methodology

Figure 3: Image composite approach in this study

The landuse dataset from Thailand's Land Development Department (LDD) in 2019 was used to generate the training dataset. First, we group LULC classes from the LDD landuse dataset into 11 categories which are applied based on The System of Environmental-Economic Accounting (SEEA) of the United Nations [30].The LULC classes are modified based on the eight traditional found in study area therefore, included: artificial surfaces (AS), herbaceous crops (HC), woody crops (WC), tree-covered areas (TC), mangroves (MG), water bodies (WB), salt fields (SF) and miscellaneous (MC), representing of geographic areas dominated by natural cover, the detailed of description publicize in Table 1. Also, we select the image of the beginning period of 2019 to generate image segmentation using the GEE platform and convert it to a vector dataset. Next, the landuse dataset and segmentation dataset in the form of vector data overlay into the dataset. Then, we perform the centroid of all polygon features representing the training location of LULC in this study. After, the centroid datasets were operated with Sentinel-2 images in 2019 to obtain the training dataset of 21 values in each centroid (five bands and three indices in three-period each).

3.2 Image Classification

We classified the LULC classes from Sentinel-2 imagery in 2019 and 2021 with the predictor variables associated with land cover for each location in the training data of 2019. After preparing the training data, we used it to predict all eight-LULC classes in Sentinel-2 imagery for the years 2019 and 2021. The land cover classification was carried out using a random forest classifier, which is one of the most commonly used classifiers for image classification based on remotely sensing data [31]. This classifier mothed has received considerable interest over the last two decades: its good handling of the outliers and noisier datasets, good performance with multi-dimension and sources datasets, and higher accuracy than other algorithms [32]. The final LULC map in 2019 and 2021 of Phetchaburi and Samut Songkhram province were produced.

Table 1: The LULC classes in this study

3.3 Accuracy Assessment

Popular measures extracted from confusion matrix reports, such as overall accuracy, producer accuracy, and user accuracy were used [34]. This process is important in determining image classification method that provides reliability and accuracy [35]. The field survey in March 2022, which is nearby image classified, was performed to collect the actual ground points covered in all classes of LULC. The equation is as follow:

Equation 1

Where:

OA or Overall Accuracy = the percentage of correctly classified pixels in the entire image or area.

Pc or Producer's Accuracy = the proportion of actual positive pixels correctly identified or classified by the model.

Nc or User's Accuracy = the proportion of actual negative pixels correctly identified or classified by the model.

Fp or False Positive Rate = the number of negative pixels classified as positive incorrectly.

Fn or False Negative Rate = the number of positive pixels incorrectly classified as negative.

3.4 Land Change Modeler (LCM)

The results of the LULC classification in 2019 and 2021 were employed in the RECLASS module of the software [36]. All values, including background values (-9999), were reclassified as class 0 to background values, class 1 to artificial surfaces, class 2 to herbaceous crops, class 3 to woody crops, class 4 to tree-covered areas, class 5 to mangroves, class 6 to water bodies, class 7 to salt fields, and class 8 to miscellaneous areas. For analyzing the change in LULC, we conducted a change analysis by comparing the differences between the LULC of 2019 and 2021.

3.5 CA-Markov Model

One of the most efficient models in current LULC modeling tools is the CA-Markov model. The CA-Markov model is a practical tool that combines

Markov and Cellular automata (CA) approaches to simulate land use changes. It is important to consider both time series and spatial data when forecasting and modeling land use changes, as this can help to better simulate the temporal and spatial patterns of these changes [37]. Markov probability is utilized to predict the future state of a system based on its present state, without considering the past states. Cellular automata, on the other hand, is a model derived from the fields of physics and biology, which uses a natural mechanism to simulate and comprehend complex behaviors [38]. We employed the integrated CA-Markov approach to predict future land use changes based on the 2019 and 2021 datasets. To achieve this, we prepared the transition probability matrix (TPM), transition probability area (TPA), and transition suitability image (TSI) using the Markov model. These outputs were then used in the CA-Markov model.

4. Results and Discussion

4.1 LULC Classification and Changing

The present study utilized Sentinel-2 time series data from 2019 and 2021 to perform LULC classification through the random forest classifier of the machine learning approach. The classification accuracy was evaluated using a confusion matrix and ground truth data obtained from field surveys. The overall accuracy of the classification was found to be 87.79%. Table 2 shows that the highest classification accuracy was obtained for water bodies (WB), salt fields (SF), and artificial surfaces (AS), respectively. It was found that separating salt fields from paddy fields or construction sites posed a challenging issue due to similarities in geometry, however, spectral classification based on the presence of even thin layers of salt was found to be useful. Furthermore, due to the limitations of the spatial resolution of the image, small artificial objects were difficult to extract, and therefore, grouping pixel technique was applied to enhance the limitations of this issue.

Overall, LULC change between 2019 and 2021, in addition to Table 3, revealed an increasing trend in human-made structures and tree plants, whereas the land for local economic functions such as HC, WC, and SF, showed a decreasing trend. Even the agricultural map, Agri-map, reported by the Land Development Department (LDD) in 2021, recommended the transformation of herbaceous plants to GI's plantation including lychee, coconut, and pomelo since 2019, following appropriate agricultural extension guidelines [39]. In contrast, Phetchaburi is renowned for its production of high-quality sea salt, key limes, palm sugar, rose apples, pineapples, and bananas, among other items [40]. However, the traditional yield has experienced a drastic decline due to a reduction in manpower in those sectors.

Table 2: Accuracy assessment of classified LULC in 2021

Table 3: The LULC changes observed between 2019 and 2021

Figure 4: Results of LULC change between 2019 and 2021

(a) gain and loss of LULC, (b) overall net change in all LULC categories, (c) artificial surface, (d) herbaceous crops, (e) woody crops, (f) tree-covered areas, (g) mangroves , (h) water bodies, (i) salt fields, and (j) miscellaneous

The Land Change Modeler (LCM) available in the software was utilized to conduct a detailed investigation of the LULC changes from 2019 to 2021. The LCM generated a grain and loss chart, depicted in Figure 4 , which represents the way LULC classes transition from one to another. The HC class was found to be decreasing from 2019 to 2021, with a conversion to WC of approximately 40 km², as seen in Figure 4(d) and 4(e).

This indicates a shift in the main economic activity, with the promotion of fruit tree plantations gradually gaining momentum in this area. Additionally, the WB areas increased by around 12.5 km², which is the highest change from SF. This provides evidence that the salt fields are being abandoned or converted into other land utilities, as shown in Figure 4(h) and 4(i). The most significant change observed was in the artificial surface (AS), which increased by almost 60 km², primarily constructed or converted from HC, as depicted in Figure 4. This LULC analysis revealed that HC experienced the most significant changes among other classes due to the continued decline of herbaceous crops' economic function in this area. In addition to its rich cultural heritage and agricultural productivity, Phetchaburi has become a popular tourist destination due to its unique topography. As depicted in Figure 5, the number of tourist accommodations has been steadily increasing since the Tourism Authority of Thailand (TAT) tourism campaign in 2017, indicating a growing demand for tourism-related activities in the region. Furthermore, the urban expansion in the area is reflected in the increasing presence of artificial surfaces, such as roads and highways, as seen in Figure 5. This expansion has led to an extension of the transportation network, facilitating the movement of people and goods within and beyond the region.

The spatial resolution of remote sensing images plays a crucial role in accurate Land Use and Land Cover (LULC) classification. However, it has been observed that a 10-meter resolution image can significantly impact the classification results, particularly in the case of objects with intricate details, such as trees and small buildings/constructions. In addition, the limited spatial resolution often leads to a mixture of spectral reflectance values, resulting in the extraction of erroneous information and confusion in near-character objects. Therefore, caution must be exercised while selecting the appropriate spatial resolution for the remote sensing image to ensure the accuracy and reliability of the LULC classification results.

4.2 The LULC Simulation in 2023

LULC change extraction is a magnificent tool for economic and urban behaviors; therefore, LULC simulation in 2023 is applied for LULC change prediction. The map of LULC change in this study from 2019 and 2021 was initial as the dependent variable to perform the MLP neural network, preparing as two classes, namely: No change and LULC change, as shown in Figure 6a. The model utilized five driving factors or variables: elevation, slope, distance to stream, distance to road, and distance to urban and village areas, as shown in Figure 6b to 6f. The dominant factors were calculated using Cramar’s V values, which determining the most dominant factor in LULC change is elevation and distance to urban and village as shown in Table 4.

Figure 5: The LULC classifications (a) 2019, and (b) 2021

Figure 6: LULC change analysis results (a) ULC Change from 2019 to 2021, (b) elevation, (c) slope, (d) distance to stream, (e) distance to road, and (f) distance to urban and village

Table 4: Sub-modules and Cramer’s V value for simulation

Table 5: The transition probability matrix for LULC change

Associated with the driving factor, a Markov chain analysis was used to produce the changing dynamic of the LULC classes. The transition probability matrix (TPM) was generated to forecast the future patterns of LULC and the results presented in Table 5.

Based on human behavior, they were unlikely to return to tree cover or other land uses once artificial surfaces were established. Therefore, it was predicted that the current artificial surface (AS) would remain as the AS in 2023 with a probability of 100%, as shown in Table 5 .

Figure 7: The LULC prediction in 2023

Table 6: The comparison of LULC change between 2019, 2021, and 2023