1. Introduction
Ports are an important node between sea and land, and port development drives land development and land use transformation in coastal zones [
1]. Additionally, the port has a close connection with the hinterland of the city where it is located. On the one hand, cities provide both a material basis and policy and economic support for port construction and expansion. A port city uses its own location advantages to establish a complete logistics system for the port. On the other hand, port-led industrial agglomerations promote infrastructure and transportation construction, further affecting the development of industry and capital to hinterland cities, which is an important reason for coastal land change [
2]. The interaction between ports and cities leads to the spatial pattern and structure of land use in port cities being different from those in general cities, and port cities exhibit a unique spatiotemporal distribution pattern from sea to land [
3,
4,
5]. The spatial pattern of land use refers to the spatial allocation of land development intensity and the order of land use transfer in this study, which reflects the spatiotemporal characteristics and gradual changes in coastal zones [
6,
7,
8].
The spatial evolution relationship between ports and cities is the focus of research on the process of port city development. Currently, the most relevant studies are focused on analyses of industrial structure, freight throughput, trade volume and other economic fields. For example, Peralta et al. (2020) presented a preliminary analysis of urban logistics spaces in Barranquilla and explored how to improve the city’s logistic system [
6]. Debrie and Raimbault (2016) studied the roles of different urban functional departments in port planning and construction by comparing ports in Venlo and Strasbourg [
7]. Mario et al. (2020) compared the development pattern of Algeciras-Gibraltar Bay in Spain and that of Osaka-Kobe Bay in Japan and studied port-city integration and the influence of large offshore engineering (reclamation works) [
8]. In addition, we cannot ignore that there are other land use transformations in coastal areas that are the product of other reasons, such as real estate speculation, reclamation works and impact on economic social policies [
9,
10,
11].
The spatial difference in land use patterns in coastal zones is very significant and requires a combination of big data for an effective analysis [
12,
13,
14,
15]. Currently, the close combination of data mining technology and geoscience studies has solved many scientific problems in geography [
12,
15]. Among them, the association rule algorithm has been widely used [
16,
17,
18,
19]. Association rule mining refers to discovering meaningful associations in a large data set to identify the correlations and laws between elements. For example, Wang (2013) searched for the optimal combination of dividing points from continuous intervals by employing genetic algorithms (GAs), in which the properties of the obtained strong association rules were treated as fitness functions to guide the algorithm iteration; then, they were used to mine the observation data from forest ecological stations to explore the relationships between forest growth and local temperature, humidity, precipitation, soil and other factors [
20]. Bembenik (2014) presented a new method of mining spatial association rules and collocations in spatial data with extended objects using Delaunay diagrams. This method does not require previous knowledge of analyzed data or the specification of any space-related input parameters and is efficient in terms of execution time [
21]. Abbes et al. (2018) proposed an efficient knowledge-based approach for vegetation monitoring using the normalized difference vegetation index (NDVI) time series combined with generated association rules to mine the relationship between climate factors and vegetation coverage in Northwestern Tunisia [
22].
At present, the application of association rule mining in geoscience research has gradually focused on combinations with specific geographical elements. Mining the correlations between geographic elements strives to make it a suitable approach for a variety of geoscience research scenarios [
17,
23]. For example, Gharbi et al. (2016) built a predictive model by mining the association rules of land use change in the process of urbanization to predict the land use change in Nancy, France [
23]. Alessandro and Tomaselli (2010) proposed and reformulated an association rule analysis (ARA) in the form of an in-depth investigation method of the spatial pattern of land cover and vegetation maps to reveal the effects of human manufacturers on the contiguous vegetation mosaics of natural and seminatural areas [
24]. Anputhas et al. (2016) structured an association rule model of land use and environmental change to predict the adjacency relations between land use changes and the environment [
25]. Ding et al. (2017 and 2019) developed a model based on the association rule algorithm to mine coastal land use sequence patterns in the sea–land direction and relied on this model to explore the spatial heterogeneity of land use in the coastal zone of Bohai Bay [
26,
27]. However, there have been few studies on the relationship between port construction and the spatial order and mutual allocation of land development intensity and land conversion mode. The impact of shipping and wharf construction on the regional development pattern and how to improve the utilization efficiency of port shorelines and to strengthen hinterland land development are problems that urgently need to be addressed in the development of coastal cities.
This study aims to reveal the sequence combination of land use patterns in the sea–land direction under different coastline change intensities in port cities, especially in port areas, and the distribution of the land use transfer pattern from sea to land. This paper selected the port areas of Qingdao and Yantai as the experimental area, and Landsat remote sensing images from 1990, 2000, 2010 and 2020 were used to extract the land use types. Then, an overlay analysis was performed to obtain land use change information. After that, according to the trend of the coastline, a sector-annular grid sample segmentation method was proposed to divide the coastal zone in the experimental area and calculate the land development intensity (LDI) level of each grid unit. The coastline and land development intensity (CLDI) model and land use transfer (LUT) model were established based on the units from sea to land to mine the spatial association rules by analyzing a decision table by means of the frequent pattern tree (FP-tree) algorithm. Finally, we compared the land development patterns and spatial association rules of the coastal zones of the two port cities and revealed the impact of port construction on the spatial pattern of the coastal zone. The results can provide a reference for exploring the coordinated development mode of ports and cities and the reasonable spatial layout of land and sea, which is conducive to promoting the sustainable development of coastal zones.
2. Materials and Methods
2.1. Study Area
This study selected the Qingdao Port area and Yantai Port area as the experimental coastal zones (
Figure 1). Qingdao is located in the south of Jiaodong Peninsula, Shandong Province, China. Qingdao features a temperate monsoon climate with an average annual precipitation of 662 mm. The north and northwest are plains, and the southwest and east are dominated by low mountains and hills. The Laoshan area in the east has a typical mountainous terrain [
28]. The ports of Qingdao, which are located in Jiaozhou Bay, mainly include Qingdao Port (the old port) and Qianwan Port. Yantai is located in the north of Jiaodong Peninsula. Qingdao also features a temperate monsoon climate, and the average annual precipitation is 525 mm. The main terrain in Yantai is low mountains and hills, and the low mountain area is located in the middle of Yantai City. The coastal plain is half-encircled by hills from west to east and faces Bohai Bay. In addition, there is a vast plain in the southeast [
29]. The ports of Yantai include Yantai Port and Yantai West Port.
The port areas of Qingdao and Yantai were both used, and an area within 25 km perpendicular to the shoreline was selected as the experimental area. This region can fully reflect the development and construction processes of ports and coastal areas and can reflect the overall expansion of building land on the basis of effectively retaining the spatial trend of the coastline. Furthermore, it can reveal the development of adjacent regions to some extent.
2.2. Research Data
In this study, remote sensing images from 1990, 2000, 2010 and 2020 were obtained from the Landsat series satellite image dataset from the USGS [
30]. The images taken in 1990, 2000 and 2010 were obtained from the Landsat-5 Thematic Mapper (TM) dataset and have a spatial resolution of 30 m, and the images taken in 2020 were obtained from the Landsat-8 Operational Land Imager (OLI) dataset and have a panchromatic band spatial resolution up to 15 m. Both were recorded in world geodetic system 1984 (WGS84) format. The radiometric calibration tool in ENVI 5.3 and the FLAASH atmospheric correction model were used to preprocess the images from various periods.
2.3. Remote Sensing Image Classification
Since the land use type information in this paper was used to calculate the LDI and represent the land use transfer pattern, only the first classification was extracted. The specific land use types included building land (BL), waters (Wa), aquaculture and salt field (AS), cultivated land (CL), forestland (FL), garden plot (GP) and other land (OL). The classification method of remote sensing images involves comprehensive index classification and visual interpretation to extract land use information from four periods (1990, 2000, 2010 and 2020).
The normalized difference water index (NDWI, Formula (1)) was used to separate Wa from the images, and then, the ASs were distinguished through visual interpretation [
31]. Second, the NDVI (Formula (2)) was used to extract vegetation. In the category of vegetation [
32], the NDVI value of FL was higher, and the NDVI value of the GPs was lower and then combined with visual interpretation to identify the boundaries of the FL, GPs and CL. The normalized difference built-up index (NDBI, Formula (3)) was established based on the reflectivity of the building land (impervious surface) in mid-infrared bands being higher than that in the near-infrared band, and the NDBI was used to extract the building land [
33].
After classifying the land use types of the Qingdao and Yantai study areas in four periods, it was necessary to revise the classification results by comparing them with each other. After that, the regions of interest (ROIs) of the land use types in different periods were selected on the Landsat TM or OLI images for precision validation, and the results were expressed by a confusion matrix. Taking the classification precision validation results of the two experimental areas in 2020 as an example, 20 samples were selected for each type of land as the ROI, and the confusion matrix is shown in
Table 1 and
Table 2. The verification results showed that the kappa coefficients were all greater than 0.8; thus, the accuracy of the interpretation met the needs of the study.
2.4. Association Rule Mining Method
Association rule mining refers to discovering meaningful associations from a large dataset, i.e., identifying frequently occurring sets of attribute values (frequent item sets) from the dataset and then using the resulting frequent item sets to describe the association rules [
34]. Whether a data point belongs to a frequent item set relies on whether the support (S) and confidence (C) exceed the support and confidence thresholds set by the operator [
35]. Yan et al. (2020) took the Port of Kuala Lumpur, Malaysia as the experimental area to establish spatial association models that reflected the relationship between the spatial pattern and allocation of land use along the sea–land gradient in the coastal zone [
36]. It was preliminarily verified that an association rule mining model could be used to explore the association rules between port shorelines and land use patterns.
This study greatly improved the model and established a new model to reflect the influence of port construction on the spatial pattern of land use in coastal zones. This improvement was mainly reflected in the following aspects: (1) According to the spatial features of the Shandong Peninsula coastal zone, this study improved the grid unit division method and made it more closely fit the coastline trend. This sample segmentation effect was more in line with the spatial features of the geographic elements. The samples were divided by constructing a sector-annular grid that matched the shape of the coastal zone. (2) The coast in this study was no longer limited to the port shoreline; rather, it included all coastline changes in the study area. We built a new CLDI model to express the correlation between the coastlines and LDI and probe how coastline changes affect land use patterns in port areas. To study the important land use transfer patterns in port areas, a new LUT model was established to express the land use transfer sequence pattern in the sea–land direction. (3) To analyze and compare the differences in the spatial utilization patterns in different port coastal zones and study the direction, distance and relevance of the port construction and land development in different regions, this research selected the two port areas of Qingdao and Yantai for a comparative study; additionally, the general law of port city expansion and the differences of the association pattern due to the different development process of ports were explored. (4) This research used the FP-tree algorithm to replace the original Apriori algorithm. Compared with the Apriori algorithm, the FP-tree algorithm based on the tree structure does not need to generate candidate frequent item sets but directly obtains frequent item sets. This structure can greatly reduce the number of times the database is traversed, thereby improving the algorithm efficiency.
2.4.1. CLDI Model
To establish the spatial association model, the study area must be divided into grid samples. The grid samples aimed to capture the sea–land distribution in coastal topography, and the samples were divided by constructing a sector-annular grid that matched the shape of the coast. First, based on the vectorization of the regional shoreline, multiple buffer zones (columns,
Figure 2) were generated from the coast to the inland areas at certain intervals, taking the two sides perpendicular to both ends of the coastline as the radius of the sector annulus and taking the coastline between the two radii as the inner arc; additionally, the outer edge that belonged to the outskirts of the buffer zones was used as the outer arc to construct a sector-annular grid that matched the shape of the coastal zone. Finally, two extension lines with radii toward the ocean were made and intersected, and the intersection of the two radii represented the center of the circle. After that, the two radii were used as the starting and ending edges, rotation segmentation (row,
Figure 2) was carried out according to certain intervals (such as 1.5°) and the vertical lines in the sea–land direction were obtained by dividing the buffer zone at each intersection to form grid units (cell,
Figure 2). This sample segmentation method not only effectively preserved the spatial trend of the shoreline but also reflected the spatial adjacency of the LDI.
After dividing grid samples, the grid samples were overlaid with land use change vector data to obtain the land use change rate of each grid unit. The LDI of the grid unit was divided into four levels using the Jenks natural breaks method: weak (rate of change < 16.93%), medium (16.93% ≤ rate of change < 38.26%), strong (38.26% ≤ rate of change < 62.88%) or extremely strong (62.88% ≤ rate of change). To reduce the data redundancy and maintain the validity of the sample information, the grid units with the same LDI level were merged into one unit (as shown in
Figure 2). After merging, the distribution of LDI in three interval periods was obtained and is shown in
Figure 3. A set of sample sequences included the units along the same direction perpendicular to the shoreline, and the sample set was used to establish the decision table of the association models.
The decision table of the CLDI association model is expressed as {Lm, DnDn ………}, where L represents the change intensity of the coastline, which is the ratio (in percentage) of the increase in the shoreline to the total length of the shoreline in the row interval; m represents the intensity level, which is 1, 2, 3 or 4, corresponding to weak (0 ≤ L < 0.75%), medium (0.75% ≤ L < 2.15%), strong (2.16% ≤ L < 7.37%) or extremely strong (7.38% ≤ L), respectively; D represents the LDI; and n represents the intensity level, which is 1, 2, 3 or 4, corresponding to weak, medium, strong or extremely strong, respectively. For example, if the coastline change in the same sample strip is strong and the corresponding LDI sequence in the landward direction is strong, medium and extremely strong, then its sample form in MATLAB is {L3, D3D2D4}, where L3 and D3D2D4 are each considered one item, so the sample {L3, D3D2D4} can be considered a 2-item set. All samples were substituted into the FP-tree algorithm in the form of 2-item sets for mining the association rules.
2.4.2. LUT Model
The LUT sequence used the same row distance as the CLDI model and took the land use transfer map patch sequence in the sea–land direction as the samples to discover the landward spatial order of the LUT types. The LUT information was obtained by superimposing the land use type maps in two interval periods. After intersecting the LUT map patches with the rows (rows,
Figure 2) perpendicular to the shoreline, the various types of map patches were sorted from sea to land. Then, excluding tiny map patches and unchanged land use types, adjacent patches with the same LUT type were merged to obtain a sequence of LUT types in Qingdao and Yantai (for example, the sequence of OL to BL > FL to GP > GP to BL), as shown in
Figure 4 and
Figure 5. The decision table established using the LUT model was expressed as {T
aT
bT
c ……}, where T represents the LUT type and a, b, c … represent the 20 LUT types, with a, b and c representing the transfer from CL to BL, FL to BL and FL to GP, respectively. Each sample was expressed as a decision table and was substituted into the FP-tree algorithm for mining the association rules. For example, the sample {FL to BL > GP to BL > FL to GP} was expressed as the item set {T
bT
hT
c}. Similarly, after the transformation into Boolean form and matrix transposition, the support and confidence were calculated based on the inner product of the matrices, and then, the important LUT sequences were extracted.
5. Conclusions
In view of the complex and dramatic changes in coastal land use patterns, this paper used remote sensing image series data to research the association between shoreline change and LDI and the LUT pattern in the sea and land directions. The spatial relationship was introduced into the association rule mining model. According to the terrain features of coastal zones, this paper designed a sample division method by constructing a sector-annular grid to determine the samples and then established spatial association models (CLDI and LUT models) that reflected the relationship between the spatial pattern and allocation of LDI in the sea–land direction, which provided reference models for the spatiotemporal analysis of the coastal zones.
Port shipping has a profound influence on the land use patterns of port cities. However, different port cities have different locational factors and development bases, so there are differences in the land use patterns along the sea–land direction. Qingdao is the core city of the region, and the interaction between the port and city is strong and has maintained a good trend of development for a long time. In the early stage, development was carried out around the Qingdao Port area and the surrounding area of the old town. In the middle stage, the key development areas moved to the Qianwan Port area. In the later stage, the transfer of port vicinity industries to the surrounding important cities and towns was the main form of land expansion. The land development intensity of the Yantai Port area was relatively high in the early and middle stages but slowed significantly in the later stage. The Strong Change Coastline -> Extremely Strong > Strong > Medium > Weak CLDI sequence and the OL → BL > CL → BL > FL → CL LUT sequence spatially coincided.
The “one city with two ports” development strategy of Qingdao and Yantai is very similar. In the early stages of development, Qingdao Port and Zhifu Port played a leading role in driving the land development of the port vicinity economic development zone. In the middle stage, land resources in the Qingdao old town adjacent to the port area were exhausted. Qianwan Port, built and put into operation in 2002, took on the functions of freight, oil and gas transport from the old port. Since Yantai planned to develop the tourism industry, Zhifu Port stopped expanding in 2013, and the newly built Yantai west port was used to divert the functions of industrial and raw mineral material imports and shipbuilding from Zhifu Port. The old port retained the passenger transport and logistics transfer functions to adapt to the development of tourism. The port construction process of the two study areas changed the spatial allocation of the LDI and LUT patterns in these coastal zones and promoted continuous adjustment and changes in the spatial structure of the port cities.
Coastal industries gradually move inward to the hinterland, along with the construction of the port area. Hence, it is necessary to plan ahead and appropriately strengthen the development of the interior. At the same time, we should note that the expansion of the city cannot be based on the occupation of FL and CL and should reasonably optimize the land use structure and spatial pattern. When the port is mature and the port vicinity industry displays an outflow trend, a new port can be planned and constructed. Furthermore, it is necessary to rationally distribute the functions of the two ports, scientifically arrange the industrial transfer from the old port, improve the land transportation network to enhance the accessibility of coastal city groups and promote coordinated development among the regions.