Introduction

The rapid growth of China’s economy, coupled with ongoing societal transformation and accelerating urbanization, reflects the nation’s increasing influence on the global stage1,2. This progress, however, has been largely driven by the expansion of urban areas. The spread of built-up land has steadily encroached upon natural environments3,4, disrupting original landscape structures5,6, altering ecological functions7,8, and degrading habitat quality9,10. These changes pose serious risks to the stability of the “production–living–ecological” (PLE) spatial framework. Consequently, the intensification of human-land conflicts has become increasingly prominent in recent years, particularly from the perspective of cultivated land, which faces the highest degree of encroachment11,12,13. As a comprehensive carrier integrating natural ecological14, economic15, and social functions16, cultivated land plays a critical role in maintaining regional landscape ecological stability, promoting agricultural development, and ensuring food security17,18,19. Its spatial distribution and landscape pattern dynamics are directly linked to the efficient utilization of agricultural resources and the coordinated development of related sectors20. In recent years, ecological research has increasingly turned its attention to cropland morphology and landscape patterns. However, most existing studies emphasize the spatial and temporal dynamics of landscape structure21, ecosystem values22, and ecosystem services23, often drawing on integrated land-use datasets that span a variety of landscape types. These efforts typically regard the structural evolution of land-use systems as the main driver of ecological change. In contrast, cropland-specific research predominantly addresses biological processes including animal behavior24, plant growth25, and soil nutrient dynamics26. Few studies treat cropland as a discrete landscape unit in its own right, particularly with respect to its spatial organization. This gap highlights the need for more targeted exploration through dedicated landscape pattern analysis.

To address these challenges, research on cropland landscapes—grounded in geographic theory—should systematically examine spatiotemporal distribution and fragmentation across three key dimensions: evaluation methods27, underlying relationships28, and driving forces29. With respect to methodology, spatial pattern evaluation can be structured around a two-level analytical framework that captures both micro-scale configuration and macro-scale structural features. For example, results derived from Morphological Spatial Pattern Analysis (MSPA) can be refined to the raster scale, enabling the acquisition of data on the fragmentation and spatial heterogeneity of cultivated land patches. This detailed analytical capability has been widely applied in identifying and studying the erosion and degradation of various landscape elements30,31. Landscape pattern indices provide a comprehensive assessment of cropland fragmentation, considering patch shape, connectivity, spatial distribution, and density32,33. Compared to MSPA, landscape pattern indices typically yield results at the regional scale, providing a more macro-level perspective. Combining these two analytical approaches ensures a more holistic and thorough evaluation. A recent study by Yang et al.34 applied MSPA and landscape pattern indices to examine cropland fragmentation dynamics in Fuqing City from 2000 to 2020. Although this study effectively captured spatiotemporal changes in cultivated land patterns, it was limited to descriptive analysis and did not explore causal links between landscape metrics or the underlying drivers of fragmentation. Furthermore, relying solely on landscape-level indices may overlook class-specific characteristics. In cropland-focused studies, class-level metrics better quantify structural attributes such as connectivity and fragmentation.

Research on the intrinsic spatial relationships of landscape elements often relies on numerical indicators such as Spearman35 and Pearson36 correlation coefficients. Spatial distribution relationships are typically examined using methods like spatial autocorrelation (univariate and bivariate)37,38 and Moran’s I analysis39. For example, Li et al.40 used bivariate spatial autocorrelation to explore trade-offs and synergies among ecosystem services, clearly visualizing spatial interdependencies across service systems. Combining correlation coefficients with spatial autocorrelation allows for the joint analysis of numerical and spatial patterns, leading to more comprehensive results.

Deeper exploration of spatial distribution patterns can be achieved through algorithms such as K-means41 and Self-Organizing Maps (SOM)40,42,43, which analyze multi-factor co-occurrence patterns to reveal spatial clustering states. Compared with conventional clustering methods like K-means, the SOM model demonstrates superior robustness. In addition, SOM results generated using the kohonen package in R can be directly visualized with sector proportion diagrams. These diagrams clearly illustrate the proportional makeup of each cluster and highlight their defining features. In studies of ecosystem service bundles, researchers including Chang et al.44, Shen et al.45, and Xia et al.46 have successfully implemented SOM algorithms, demonstrating this model’s broad applicability in landscape ecology research.

Regarding studies on driving factors, human activities have been identified as critical determinants influencing spatial distribution patterns and morphological characteristics47. The complex interplay of anthropogenic activities introduces uncertainties into the intervention levels and status of cultivated land landscape patterns. A comprehensive selection of drivers typically encompasses natural, economic, and social dimensions. Current methodological approaches primarily employ models including GeoDetector48,49, Geographically Weighted Regression (GWR)39, Geographically Temporally Weighted Regression (GTWR)50, and Multiscale Geographically Weighted Regression (MGWR)51,52,53. Unlike GWR, which applies a fixed bandwidth, MGWR assigns variable bandwidths to different factors, improving model accuracy. Mandal et al.54 showed that MGWR outperformed GWR across all metrics in analyzing spatial heterogeneity in land surface temperature. This three-tiered analytical framework enables a comprehensive understanding of spatial dynamics, interrelationships, and the drivers of cultivated land patterns following erosion.

Growing concerns about food security, alongside changing global economic dynamics, have refocused strategic attention on agricultural resilience in China55,56. This study focuses on Shandong Province, a major agricultural and economic region under dual pressures from urban expansion and farmland conservation. Our analysis is structured around three core methodological components:

  1. 1.

    A multiscale fragmentation analysis combining MSPA and landscape metrics to track cropland subdivision across Shandong (1990–2023), offering a micro-to-macro perspective on fragmentation dynamics;

  2. 2.

    Analysis of pattern–cluster relationships through bivariate spatial autocorrelation, Pearson correlation, and SOM clustering to clarify inter-metric relationships and define cluster structures within the landscape metric set;

  3. 3.

    Investigation of fragmentation drivers using MGWR to untangle urbanization-driven pathways, producing spatially detailed insights to inform county-level land use policy.

Materials and methods

Overview of the study area and data sources

Overview of the study area

Shandong Province is located in the eastern coastal region of China, within the lower reaches of the Yellow River. It borders the Bohai Sea to the northeast and the Yellow Sea to the southeast. The inland region is adjacent to Jiangsu, Anhui, Henan, and Hebei provinces. Geographically, it lies between 114.81°E–122.71°E and 34.38°N–38.40°N (Fig. 1) (Projection coordinates: WGS_1984_UTM_Zone_50N). Shandong Province comprises 16 prefecture-level divisions (136 county-level units) spanning 15.67 × 103 km2, with Jinan as the capital and Tai’an, home to Mount Tai, the most famous of the Five Great Mountains, situated within its borders. Shandong experiences a warm temperate monsoon climate, with an average annual temperature of approximately 13 °C and annual precipitation ranging from 600 to 750 mm. By the end of 2023, the province had a permanent population of approximately 101.23 million people, with a GDP of 9.21 trillion RMB. As a major agricultural province, Shandong is known for its high and high-quality yields of both food and cash crops. The province ranks third in the nation in terms of the planted area and production of food crops, including wheat and maize. Major cash crops include peanuts and cotton. Additionally, Shandong serves as a key production and supply base for vegetables across the country, reinforcing its crucial role in national agriculture (Statistical year data of Shandong Province and Baidu Encyclopaedia data).

Fig. 1
figure 1

Geographic location map of the study area (created using ArcGIS 10.8, https://www.esri.com).

Full size image

Data sources and research framework

The land use data used in this study were obtained from the Wuhan University team, which released Chinese surface coverage data for four time periods (1990, 2000, 2010, and 2023) with a spatial resolution of 30 m, resampled to 100 m. The DEM data were sourced from the SRTMDEM 90 M dataset (Geospatial Data Cloud Platform), and the slope data were derived from the DEM calculations. NDVI, NPP, and nighttime light coefficient data were obtained from the Resources and Environmental Science Data Platform. Road density data were based on OpenStreetMap road length data and the area ratio of districts/counties. Urbanization rates were calculated using the ratio of urban area to district/county area. Population, GDP, and crop yield data were sourced from the Statistical Yearbooks of various cities in Shandong Province. The distance to rural settlements was derived by applying Euclidean distance processing to the rural settlement data from the Chinese Academy of Sciences Land Use and Land Cover Change Database.

This study is divided into three research phases (Fig. 2): (1) Fundamental exploration of landscape patterns: Utilizing MSPA and landscape pattern indices, a comprehensive analysis of Shandong Province’s cropland from 1990 to 2023 was conducted. Changes in the area and distribution of landscape elements such as core zones, islets, and perforations were examined to assess the degree of cropland fragmentation and its evolution during urbanization. Additionally, six landscape pattern indices were selected to analyze cropland aggregation, connectivity, and geometric regularity across from multiple perspectives 136 districts/counties. (2) Investigation of spatial relationships: The relationships among landscape pattern indices were explored using BSA and Pearson correlation analysis. Subsequently, SOM clustering was employed to classify cropland in Shandong Province, providing insights for tailoring policy recommendations to cropland with varying characteristics. (3) In-depth analysis of driving factors: Based on the MGWR model, 11 indicators were selected to investigate the factors influencing cropland landscape patterns.

Fig. 2
figure 2

Research framework diagram (created using GuidosToolbox 3.0, https://forest.jrc.ec.europa.eu/en/activities/lpa/gtb/; R 4.3.3, https://cran.r-project.org/; Origin 2022, http://www.originlab.com/; FRAGSTATS v4.2.1, https://www.fragstats.org/; GeoDa 1.20, https://geodacenter.github.io/; ArcGIS 10.8, https://www.esri.com; MGWR 2.2, https://github.com/pysal/mgwr; Microsoft Office 2021, https://www.microsoft.com/zh-cn/download/office).

Full size image

Research methodology

MSPA

MSPA is a method based on mathematical principles that involves operations such as erosion, dilation, and other opening and closing calculations to identify and segment elements in binary raster images. After identifying significant foreground patches, further operations like erosion and expansion are performed to reconstruct the spatial structure, resulting in seven types of landscape pattern elements57,58,59 (Table 1). In this study, agricultural land patches were extracted as foregrounds for analysis, with the edge width parameter set to 1 pixel, to explore the evolution of agricultural landscape patterns during the urbanization process.

Table 1 Ecological significance of the MSPA7 landscape elements.
Full size table

The single dynamic degree of change represents the trend of development and change of the study element over a specific period. This study analyzes the fragmentation of agricultural land based on the dynamic degree of change of seven MSPA landscape types. The calculation formula is as follows:

$${\text{V}} = \frac{{L_{m} - L_{n} }}{{L_{n} \times T}} \times 100\%$$
(1)

\(L_{m}\),\(L_{n}\) represent the quantities of the element at time points \(m\),\(n\), respectively; \(T\) is the time interval between \(m\),\(n\), and \({\text{V}}\) denotes the dynamic degree of the element. A value greater than 0 indicates an increase.

Landscape pattern index

This study selects six landscape pattern indices from the type scale to conduct a detailed analysis of the changes in the cultivated land landscape pattern due to urban expansion from 1990 to 2023. The combination analysis of landscape pattern indices over a long time series can, to some extent, reflect the evolution of cultivated land spatial distribution and composition in the study area67,68. Using districts/counties as the research units, the landscape pattern indices of cultivated land in 136 districts/counties of Shandong Province were calculated. The calculation formulas are as follows:

$${\text{CONNECT}} = \frac{{\sum\nolimits_{j = k}^{n} {c_{ijk} } }}{{\frac{{n_{i} (n_{i} - 1)}}{2}}} \times 100$$
(2)

CONNECT represents the functional connectivity between cultivated land patches. The index is calculated as the ratio of the number of connectivity path nodes within a critical distance between cultivated patches to the total number of connectivity path nodes in the region. A higher value indicates a greater degree of patch connectivity21.

$${\text{PLAND}} = \frac{{\sum\limits_{j = 1}^{n} {a{}_{ij}} }}{A} \times 100$$
(3)

PLAND represents the proportion of cultivated land area within a region. Changes in PLAND values at different time points in the same area can intuitively reflect the extent of cultivated land loss due to urban construction and land expansion. A decrease in PLAND indicates a reduction in the area of cultivated land69,70.

$${\text{PD}} = \frac{NP}{A}$$
(4)

PD represents the patch density of cultivated land. Over a long time series, changes in PD values within the same district or county can reflect the degree of fragmentation of cultivated land patches. An increase in PD values indicates a deeper fragmentation of cultivated land in the region70,71.

$${\text{PAFRAC}} = \frac{2}{{\frac{{N\sum\limits_{{{\text{i}} = 1}}^{{\text{m}}} {\sum\limits_{{{\text{j}} = 1}}^{{\text{n}}} {\left( {{\text{lnp}}_{{{\text{ij}}}} \times \ln a_{{{\text{ij}}}} } \right)} } - \sum\limits_{{{\text{i}} = 1}}^{{\text{m}}} {\sum\limits_{{{\text{j}} = 1}}^{{\text{n}}} {\ln } } {\text{p}}_{{{\text{ij}}}} \times \sum\limits_{i = 1}^{m} {\sum\limits_{j = 1}^{n} {\ln } } a_{ij} }}{{N\sum\limits_{i = 1}^{m} {\sum\limits_{j = 1}^{n} {\ln } } p_{ij}^{2} - \sum\limits_{i = 1}^{m} {\sum\limits_{j = 1}^{n} {\ln } } p_{ij} }}\quad }}$$
(5)

PAFRAC represents the degree of shape complexity of cultivated land patches, with values ranging from. A PAFRAC value closer to 1 indicates a simpler, more regular shape of the patches, while a value approaching 2 suggests more complex and irregular patch shapes. The PAFRAC value reflects the extent of interference or erosion affecting the cultivated land, with higher values indicating greater levels of disturbance72.

$${\text{COHESION}} = \left[ {1 - \frac{{\sum\limits_{j = i}^{m} {p_{ij} } }}{{\sum\limits_{j = 1}^{m} {p_{ij} \sqrt {a_{ij} } } }}} \right]\left[ {1 - \frac{1}{\sqrt Z }} \right]^{ - 1} \times 100$$
(6)

COHESION is the cohesion index, which measures the connectivity of cultivated land patches. The index ranges from 0 to 100, with lower values indicating a higher degree of fragmentation due to erosion and poorer physical connectivity. Conversely, higher values indicate greater aggregation and connectivity among the patches70,73.

$${\text{AI}} = \left[ {\frac{{{\text{g}}_{{{\text{ii}}}} }}{{\max \to {\text{g}}_{{{\text{ii}}}} }}} \right](100)$$
(7)

Aggregation Index represents the degree of aggregation between different cultivated land patches or regions, with a value range of [0,100]. A higher AI value indicates a higher density of cultivated land patches, whereas a lower value indicates greater dispersion of these patches74.

Bivariate spatial autocorrelation

The landscape pattern index distribution state and the association and dependence pattern between them were explored based on bivariate spatial autocorrelation (BSA) analysis with the following formulae:

$${\text{Moran}}\prime {\text{s I(Bivariate)}} = \frac{{\sum\limits_{i = 1}^{n} {\sum\limits_{j = 1}^{n} {W_{ij} } } \left( {x_{i} - \overline{x}} \right)\left( {y_{j} - \overline{y}} \right)}}{{S^{2} \sum\limits_{i = 1}^{n} {\sum\limits_{j = 1}^{n} {W_{ij} } } }}$$
(8)
$${\text{I}}_{{\text{i}}} = Z_{i} \sum\limits_{j = 1}^{n} {W_{ij} } Z_{j}$$
(9)

(8) and (9) represent global and local spatial autocorrelation formulas, respectively. In this study: \(n\) is the number of samples in districts/counties of Shandong Province; \(x_{i}\) and \(y_{j}\) represent the values of landscape pattern indexes represented by \(X\) and \(Y\) at districts and counties \(i\) and \(j\), respectively; \(\overline{x}\) and \(\overline{y}\) are the mean values; \(W_{ij}\) is the spatial weighting; \(S^{2}\) is the variance of the landscape pattern index of the variable; and \(Z_{i}\) is the deviation of the landscape pattern indexes from the mean value.

Pearson correlation coefficient

The Pearson correlation coefficient, proposed by Karl Pearson based on previous research, is a widely used method for measuring the linear correlation between two variables, X and Y75. It has been extensively applied in fields such as mathematics and ecology:

$$r = \frac{1}{n - 1}\sum\limits_{i = 1}^{n} {\left( {\frac{{X_{i} - \overline{X}}}{{\sigma_{X} }}} \right)} \left( {\frac{{Y_{i} - \overline{Y}}}{{\sigma_{Y} }}} \right)$$
(10)

\(r\) is the correlation coefficient, \(\overline{X}\),\(\overline{Y}\) is the mean; \(\sigma_{X}\),\(\sigma_{Y}\) is the standard deviation; \(n\) is the number of samples. In this study \(r\) was used to analyze the correlation between 6 indices of cropland landscape pattern.

SOM clustering algorithm

SOM algorithm, a type of neural network model, is capable of reducing the dimensionality of high-dimensional data and performing visual clustering analysis76. Its calculation process is as follows:

$$d({\text{x,w}}_{{\text{i}}} ) = \sqrt {\sum\limits_{j = 1}^{n} {(x_{j} - w_{ij} )^{2} } }$$
(11)
$${\text{w}}_{{\text{i}}} (t + 1) = {\text{w}}_{{\text{i}}} (t) + \alpha (t) \cdot hci(t) \cdot ({\text{x}} - {\text{w}}_{{\text{i}}} (t))$$
(12)

\(d({\text{x,w}}_{{\text{i}}} )\) represents the Euclidean distance; \({\text{w}}_{{\text{i}}}\) is the weight vector of neuron \(i\); and \({\text{w}}_{{\text{i}}} (t + 1)\) denotes the updated weight. Iterative computations are conducted based on the formula until the network converges.

In this study, the SOM algorithm was implemented in R to cluster six cropland landscape pattern indices. The optimal number of clusters was determined to be six by identifying the inflection points in both the within-cluster sum of squares and the between-cluster sum of squares.

MGWR

The MGWR model was introduced by the Fotheringham team in 2017 as an improvement upon the GWR model. Unlike the GWR model, which uses a fixed bandwidth, MGWR assigns variable bandwidths to different predictors, thereby enhancing the accuracy of regression results77. The GWR calculation formula is as follows:

$${\text{y}}_{i} = \beta_{0} (u_{i} ,v_{i} ) + \sum\limits_{k = 1}^{m} {\beta_{k} (u_{i} ,v_{i} )x_{ik} + \varepsilon_{i} }$$
(13)

\(x_{ik}\) and \({\text{y}}_{i}\) represent the independent and dependent variables at location \(i\), respectively, while \(\beta_{k} (u_{i} ,v_{i} )\) denotes the regression coefficients. In this study, based on insights from previous research and considering the specific characteristics of the study area, an analysis was conducted using 2023 data. Eleven indicators were initially selected as potential driving factors of cropland landscape pattern heterogeneity at the district/county level: urbanization rate, DEM, slope, NDVI, road density, distance to rural settlements, NPP, nighttime light coefficient, GDP, crop yield per unit area, and per capita cropland area. These indicators were used in a MGWR analysis to assess spatial heterogeneity.

Results

Morphological spatial pattern analysis of cultivated land

Based on the cropland distribution data from four periods (1990–2023), MSPA was conducted to classify cropland into seven types of morphologically distinct landscape patterns (Fig. 3). As illustrated in detail panels I and II, the spatial extent of green core areas has progressively declined over time. This contraction is especially pronounced in several key regions of Shandong Province, including Jinan, Zibo, Tai’an, and Weifang in the central region; Yantai and Qingdao on the northeastern peninsula; and Linyi in the southeast. In these cities, rapid economic growth driven by industrialization, technological advancement, and tourism development has intensified the expansion of construction land, thereby exacerbating the encroachment on arable land resources.

Fig. 3
figure 3

MSPA and Core Area Delineation of Cultivated Land in Shandong Province from 1990 to 2023 (created using GuidosToolbox 3.0, https://forest.jrc.ec.europa.eu/en/activities/lpa/gtb/; ArcGIS 10.8, https://www.esri.com).

Full size image

To further evaluate the extent and status of core area changes, a classification scheme was applied based on natural breaks in ArcGIS. Thresholds of 100, 500, 1000, 2000, and 5000 were used to divide core areas into six hierarchical levels. Using the detail maps in Panel III, we examined changes in the spatial distribution of core areas at each level from 1990 to 2023. In 1990, Level 1 core areas were distributed in a ring-like pattern around the periphery of Shandong Province. Over time, this distribution diminished, and by 2010, the ring structure fragmented into three distinct regions: the northwest (Liaocheng, Dezhou, and Binzhou), the southwest (Heze and Jining), and the southeast (Weifang, Rizhao, and Linyi). By 2023, the extent of Level 1 core areas had further contracted, with remaining fragments concentrated in the northwest and central-eastern parts of the province. This pattern of decline reveals a clear spatial correlation: regions with higher levels of economic development exhibit lower proportions of Level 1 core areas. The gradual disintegration of these areas—initially dominant in central and northeastern Shandong—reflects the dynamics of urban expansion and the ongoing encroachment upon arable land in economically advanced zones.

The analysis of the quantity, area, and proportion of the six levels of core areas (Table 2) reveals notable trends. The number of Level 1 core areas remained relatively stable; however, their area proportion declined significantly from 82.44% in 1990 to 30.28% in 2023, indicating a consistent downward trend over time. In contrast, Levels 2–5 core areas showed a gradual increase in both number and proportion. This shift can be attributed to the erosion and fragmentation of Level 1 core areas, which caused them to degrade into lower-level core areas. Level 6 core areas exhibited a marked increase in quantity, area, and proportion, with the number of core areas rising from 29,225 in 1990 to 51,529 in 2023, far exceeding that of any other level. Overall, from 1990 to 2023, cropland in Shandong Province experienced significant erosion and an increasing degree of fragmentation.

Table 2 Quantities, Areas, and Proportions of Core Areas at Different Levels. Unit: km2.
Full size table

An analysis of the dynamic changes in the area of the seven morphological spatial pattern landscape types from 1990 to 2023 (Fig. 4) reveals a clear trend of accelerated reduction in core area size. The annual rate of decrease in core area intensified from −0.71% during 1990–2000 to −1.14% during 2010–2023. This contraction of core areas also led to a reduction in the transitional perforation zones between core and non-core areas, as well as the internally connected islet zones. The rate of decrease in perforation zones outpaced that of islet zones. Additionally, as core areas underwent erosion and fragmentation, their numbers increased, resulting in a growth in the area of connecting bridges between core areas. During the three time intervals of 1990–2000, 2000–2010, and 2010–2023, the dynamic acceleration of islet area size initially increased and then decreased, reflecting a continuous process of cropland fragmentation in Shandong Province, though with a decreasing intensity over time. The dynamic rates of edge area were 4.42%, 2.36%, and 2.08% for the three periods, respectively, showing a trend of increasing and then decelerating growth. This increase in edge area can be attributed to the advancing urbanization, which led to the encroachment of large areas of cropland and a gradual intensification of fragmentation, along with an increase in the irregularity of cropland edges. However, in recent years, factors such as slower urban expansion and the strengthening of cropland protection policies have reduced the level of cropland erosion, resulting in a deceleration in the increase of edge area.

Fig. 4
figure 4

Dynamics of the Area Changes for the Seven Landscape Classes (created using Origin 2022, http://www.originlab.com/).

Full size image

Analysis of the spatial differentiation and evolution of cultivated land landscape pattern indices

Using Fragstats software, the spatial distribution of six landscape pattern indices across districts/counties in Shandong Province was calculated. Based on data from 1990 to 2023, the natural breaks method was applied to adjust the threshold values, dividing the indices into five categories (Fig. 5). For the PLAND, higher category levels represent a greater proportion of cropland area within a region. The results show that in 1990, with the exception of Jinan in the central region and Yantai in the northeast, most areas exhibited high cropland proportions. Level 5 zones were predominantly arranged in a ring surrounding the provincial periphery. Over time, this ring structure gradually disintegrated. Between 1990 and 2023, the proportion of districts/counties assigned to Level 5 declined from 66.91 to 30.88%. This contraction was particularly evident in central Shandong (Jinan, Zibo, and Weifang) and along the eastern coast (Qingdao, Yantai, and Weihai). These areas have experienced rapid economic growth, where urban expansion has increasingly encroached upon cultivated land, leading to substantial reductions in cropland coverage.

Fig. 5
figure 5

Spatial distribution and proportion of landscape pattern indices for cropland in Shandong Province from 1990 to 2023 (created using ArcGIS 10.8, https://www.esri.com; FRAGSTATS v4.2.1, https://www.fragstats.org/; Origin 2022, http://www.originlab.com/).

Full size image

Overall, the PD index in 1990 exhibited a spatial pattern characterized by higher values in the central region (notably around the provincial capital, Jinan) and the northeast (including coastal port cities such as Qingdao and Yantai), and lower values in peripheral areas. This pattern reflects the earlier onset of economic development in these urban centers, where built-up land had already encroached upon cropland at the beginning of the study period, leading to the fragmentation of agricultural parcels and thus a higher patch density. As economic development progressed across other parts of Shandong, urban expansion increasingly contributed to cropland fragmentation. By 2023, the proportion of districts/counties classified as Level 5 had increased from 9.56% to 20.59%, primarily concentrated in the central, northeastern, and southern parts of the province (Fig. 5). In contrast, areas categorized as Level 1 declined sharply from 45.58 to 5.88%. These changes indicate that cropland previously organized in large, contiguous tracts has been subdivided into smaller, more fragmented units under pressure from urban growth.

The AI index, which measures the degree of clustering versus dispersion among cropland patches, further corroborates this trend. Higher AI values indicate more compact and cohesive cropland distributions. In 1990, high aggregation was observed in northern Shandong (Binzhou, Dongying), western areas (Liaocheng, Heze), and parts of the east (Qingdao) and southeast (Rizhao). By 2023, however, AI values had declined markedly across the study area (Fig. 5). The proportion of districts/counties in the highest AI category dropped from 31.62% to 0%, while Levels 1 and 2 rose steadily. This decline was especially pronounced in the central and eastern coastal regions, where rapid urbanization has led to increased fragmentation and spatial disintegration of cropland patches.

The spatial distribution of the CONNECT index across Shandong generally exhibits higher values in the west and lower values in the east. This suggests that in the eastern part of the study area, cropland patches are more functionally interconnected, facilitating more efficient ecological flows of energy, materials, and information. Between 1990 and 2023, cropland connectivity across the region declined markedly due to ongoing urban encroachment. The proportion of districts and counties with high CONNECT index values (Levels 4 and 5) decreased, with Level 5 dropping from 11.76% to just 0.74%. These reductions were most prominent in central-northern cities such as Binzhou, Zibo, and Weifang, as well as Heze in the southwest. Meanwhile, the proportions of low-connectivity areas (Levels 1 and 2) increased by 18.38% and 10.28%, respectively (Fig. 5).

The COHESION index, which measures the physical integrity and fragmentation of cropland, initially exhibited low values in central Shandong and high values around its periphery. In 1990, areas classified as Level 5—formed a ring-like structure surrounding the province. However, this cohesive ring gradually disintegrated over time, with the Level 5 proportion declining from 74.26 to 47.79% by 2023. The most notable reductions occurred in central areas such as Jinan, Zibo, and Weifang, as well as eastern and southeastern coastal cities including Yantai, Weihai, Linyi, Qingdao, and Rizhao (Fig. 5).

The PAFRAC reflects the geometric complexity of cropland patches, with lower values indicating more regular, compact shapes. Across Shandong Province, higher PAFRAC values were observed in the central, northern, and northeastern regions—areas encompassing the provincial capital and rapidly developing coastal cities. These regions experienced intensified urban expansion, resulting in more irregular cropland configurations compared to other parts of the province. However, a temporal analysis from 1990 to 2023 reveals an overall decline in PAFRAC values, indicating a trend toward simpler, more regular cropland shapes. The proportion of areas classified as Levels 4 and 5 dropped significantly, with Level 5 declining from 19.85 to 5.88%. These changes were most pronounced in central cities such as Zibo and Weifang, southern areas including Zaozhuang and Linyi, and northern regions such as Dongying and Binzhou (Fig. 5). This suggests that cropland in these areas has shifted from initially fragmented and irregular forms toward more consolidated and geometrically regular configurations. A closer analysis reveals two primary drivers behind these spatial and temporal trends. First, intensified conversion of cropland—reflected by decreasing PLAND values in the affected areas—indicates substantial land encroachment for urban development. Notably, recent land-use policies guided by sustainability principles have restricted the occupation of contiguous, high-quality permanent farmland. As a result, irregular cropland patches were preferentially targeted for development, leading to a relative reduction in the proportion of geometrically complex plots. Second, the implementation of cropland protection policies in recent years has promoted land consolidation initiatives. These efforts have effectively regularized previously fragmented and irregular cropland parcels, contributing to the observed decrease in PAFRAC values.

A comprehensive analysis of the spatial heterogeneity of cropland in the study area was conducted by examining the spatial distribution and changes of six landscape pattern indices. Between 1990 and 2023, due to the advancement of urbanization, the expansion of built-up areas encroached upon large portions of cropland. This resulted in a general decline in cropland area, an increase in patch fragmentation, a decrease in connectivity, and a rise in patch dispersion across Shandong Province. These changes reflect ecological degradation of cropland, hindering its efficient use. From a macro-scale perspective, these changes were concentrated in municipalities with faster economic development. At a finer scale, they predominantly occurred within core urban districts, highlighting the direct impact of urban expansion on cropland spatial distribution—an effect particularly pronounced in rapidly developing regions. Notably, the degree of patch shape irregularity has improved over time, reflecting the initial effectiveness of cropland protection policies and suggesting that these efforts should be further strengthened. Overall, the western part of Shandong Province has a higher proportion of cropland area, with more contiguous patches and a higher degree of aggregation, making it suitable for large-scale mechanized farming.

Bivariate spatial autocorrelation analysis of arable land landscape pattern indices

A BSA analysis of cropland landscape pattern indices from 1990 to 2023 was conducted to obtain the Moran’s I, Z-values (Table 3), and spatial dependency relationships (Fig. 6). The p-values indicate that, with the exception of the CONNECT index, all other five indices exhibited statistically significant pairwise correlations throughout the study period, and their spatial correlation patterns remained relatively consistent. In contrast, the CONNECT index demonstrated significant spatial correlation only with the PAFRAC index in 1990 and 2000, and with the COHESION index in 2023. Among the indices, AI-PLAND exhibited the highest mean I value over the four years, at 0.6516, with a corresponding mean Z-value of 13.9541, indicating a strong positive spatial correlation. High-high clusters were observed in the southwestern and northwestern regions of Shandong, while low-low clusters were found in the central-western and eastern regions. In contrast, AI-PD demonstrated the closest mean I value to −1, at −0.6420, with a corresponding mean Z-value of −13.8030, reflecting a significant negative spatial correlation. High-low clusters were concentrated in the southwestern and northwestern regions, while low–high clusters were evident in the central-western region.

Table 3 Moran’s I and Z obtained from BSA analyses of landscape pattern indices.
Full size table
Fig. 6
figure 6

BSA analysis of arable land landscape pattern indices (created using GeoDa 1.20, https://geodacenter.github.io/; Microsoft Office 2021, https://www.microsoft.com/zh-cn/download/office).

Full size image

Overall, significant spatial clustering results for other indices correlated with PLAND and PD were primarily distributed in the western regions of Shandong, including Heze, Liaocheng, Jining, and Dezhou, as well as the central areas of Jinan and Zibo, and coastal areas in southeastern Qingdao. For the spatial autocorrelation analysis of other indices with PAFRAC, significant clustering was concentrated in the southwestern areas of Heze and Jining, the central Taishan mountainous regions, and the northern areas of Binzhou and Dongying. Regarding the analysis of COHESION, significant clusters were mainly distributed in the western and southeastern parts of Shandong Province.

Based on the analysis of the number of significant clustering results (Fig. 7), the AI-COHESION correlation exhibited the highest number of "high-high" clustered districts/counties, with a four-year average of 27. The PAFRAC-PD correlation showed the largest number of "low-low" clustered districts/counties, with an average of 22. For "high-low" clustering, the largest number of districts/counties was observed in the COHESION-PD and AI-PD correlations, with an average of approximately 31. Similarly, the highest number of "low–high" clustered districts/counties was found in the PD-PLAND correlation, with an average of about 25.

Fig. 7
figure 7

Proportional Distribution of Results by Bivariate Spatial Autocorrelation Types (created using GeoDa 1.20, https://geodacenter.github.io/; Origin 2022, http://www.originlab.com/; Microsoft Office 2021, https://www.microsoft.com/zh-cn/download/office).

Full size image

Correlation analysis of cultivated land landscape pattern indices

Based on the Pearson correlation coefficients between six landscape pattern indices for cropland in Shandong Province from 1990 to 2023, negative correlations were observed for PLAND-PD, PLAND-PAFRAC, PD-COHESION, PD-AI, PAFRAC-AI, and CONNECT-COHESION across all four periods, with p ≤ 0.01. Among these, the PLAND-PD correlation had the highest negative coefficient, with an average value of -0.9325 across all periods, indicating that regions with higher PLAND values tend to have lower PD values. The second strongest negative correlation was observed between PD-COHESION, with an average coefficient of -0.8625. In contrast, positive correlations were found for PLAND-AI, COHESION-AI, PLAND-COHESION, and PD-PAFRAC, indicating that the spatial distributions of these indices tend to vary in tandem, with both increasing or decreasing simultaneously, and p ≤ 0.01. Among these, PLAND-AI exhibited the highest positive correlation coefficient, with an average value of 0.9225 across all periods. The average Pearson correlation coefficients for the other three pairs were 0.8275, 0.6625, and 0.4925, respectively (Fig. 8).

Fig. 8
figure 8

Pearson Correlation Coefficient Plot of Landscape Pattern Indices (created using Origin 2022, http://www.originlab.com/).

Full size image

Analysis of clustering results for cultivated land landscape patterns

Cluster analysis using the SOM method identified six distinct landscape pattern index clusters, labeled A–F (Fig. 9). Cluster A is characterized by high values for COHESION, AI, and PAFRAC, and medium values for PLAND and PD. It is marked by relatively aggregated and irregular cropland patches with a small area proportion. In 1990, Cluster A was predominantly concentrated in central and northeastern Shandong. Over time, their spatial distribution gradually expanded to northern, southern, and southeastern coastal regions of the province. By 2023, the proportion of districts/counties containing these clusters increased from 13.2 to 31.6%, representing the most significant growth rate among all categories, with most newly incorporated areas transitioning from previous Cluster D. This spatial evolution reflects three concurrent agricultural transformations: decreased proportional area of cultivated land, increased cultivation density, and enhanced spatial fragmentation within remaining agricultural zones. These changes have been principally driven by urban expansion emanating from major economic hubs such as Jinan and Qingdao. The spatial propagation of urbanization has stimulated regional development while concurrently escalating land demand for secondary and tertiary industries, resulting in substantial conversion of agricultural land. This territorial reorganization has fundamentally altered traditional agricultural landscape configurations. For regions dominated by Cluster A, strategic land-use planning must balance manufacturing growth with agricultural preservation. Implementing integrated development policies could potentially achieve synergistic advancement of both industrial and agricultural sectors. Cluster B is characterized by high values for PD, CONNECT, and PAFRAC, with high cropland patch density, good connectivity, but irregularity. This cluster has the smallest proportion and is mainly found in the central urban areas of Qingdao and Jinan, where the cropland area is minimal. The high patch density in these areas is due to the small size of the regions, which are economically highly developed and unsuitable for agricultural production. Cluster E is characterized by high values for PAFRAC, PD, and COHESION, and medium values for AI and PLAND. It is marked by small cropland area, highly irregular patches, high patch density, and greater aggregation. This cluster is primarily located in the central urban areas of Jinan, Qingdao, Zibo, Yantai, and Weifang, and is similarly unsuitable for agricultural development. During the 1990–2023 period, Clusters B and E exhibited moderate growth with relatively stable trajectories. These clusters demonstrate highly concentrated economic development patterns, suggesting that future planning in these regions may prioritize non-agricultural sectors.

Fig. 9
figure 9

Landscape pattern index clustering types, spatial distribution of clustering results, quantitative changes and transfer matrix chords (created using ArcGIS 10.8, https://www.esri.com; R 4.3.3, https://cran.r-project.org/; Origin 2022, http://www.originlab.com/; Microsoft Office 2021, https://www.microsoft.com/zh-cn/download/office).

Full size image

Clusters C, D, and F share fundamental agricultural characteristics: substantial cultivated land proportions, high spatial density, and superior landscape connectivity—attributes particularly conducive to crop cultivation. Their suitability ranking follows C > F > D based on comprehensive evaluations. Geographically, these clusters predominantly occupy western, northwestern, central-eastern, and southeastern Shandong. Cluster C is specifically characterized by high PLAND, AI, and COHESION metrics, coupled with moderate CONNECT values, indicating superior agricultural land aggregation and connectivity. Notably, Cluster C experienced the most dramatic proportional decline, decreasing from 26.5 to 3.7% during the study period. Its spatial distribution gradually contracted from western, central-northern, and southern Shandong in 1990 to fragmented distributions in southern and northern regions by 2023. Analysis of transition matrices reveals that diminished Cluster C areas transitioned predominantly into Clusters F and D. While these converted lands remain agriculturally viable, their reduced suitability ranking suggests compromised cultivation potential compared to original Cluster C configurations. Cluster D is characterized by high values for PLAND, AI, and COHESION, and a medium value for PAFRAC. It features a high cropland area proportion, a relatively aggregated spatial distribution, but irregular cropland patches. Cluster F is characterized by high values for PLAND, AI, and COHESION, and is distinguished by a high cropland area proportion and a relatively aggregated spatial distribution. From 1990 to 2023, the proportion of Cluster D decreased from 36.8 to 25.7%, while Cluster F increased from 17.6 to 27.9%. This conversion process demonstrates two distinct transformation pathways: Cluster D primarily transitioned into Clusters F and A, representing contrasting agricultural suitability outcomes. The D-A conversion indicates erosion-induced degradation of cultivated land with diminished agricultural suitability, whereas the D-F transformation reflects improved cultivation potential through land optimization. These transitional patterns became particularly evident during 2010–2023, coinciding with the implementation of cultivated land protection policies around 2010. Systematic field consolidation and land rehabilitation initiatives—including land parcel regularization and agricultural plot mergers—substantially enhanced cultivation suitability in converted areas. Mechanistic analysis confirms that policy-driven landscape modifications directly contributed to the accelerated suitability improvements observed in D-F transitions during this phase.

The chord diagram of the transition matrix reveals the following patterns: 1990–2000: The top three transitions in terms of area were D–F (16,310.46 km2), F–D (11,546.01 km2), and C–F (6,468.60 km2). While the transition from D to F indicates improved conditions, the transitions from F to D and C to F signify increased fragmentation and reduced connectivity of cultivated land. However, these areas remained suitable for farming. 2000–2010: The leading transitions were F–D (10,813.00 km2), D–F (9972.20 km2), and D–A (8340.09 km2). The transition from D to A represents a reduction in cultivated land area and increased fragmentation, rendering the land unsuitable for farming. 2010–2023: The top three transitions were D–F (16,960.07 km2), D–A (14,266.98 km2), and F-D (12,301.91 km2). On the one hand, some cultivated land became more consolidated and regular in shape (D–F). On the other hand, there was a further reduction in cultivated land area and increased fragmentation (D–A). This scenario is primarily driven by urban expansion encroaching on agricultural land and intensifying fragmentation. Simultaneously, recent government policies aimed at farmland protection and agricultural development have started to show positive effects. From 1990 to 2023 as a whole, the various contradictions existing in the spatial pattern of arable land in Shandong gradually deepened.

Analysis of driving factors for the distribution of cultivated land landscape patterns

The significant driving factors for each landscape pattern index vary. Initially, MGWR analysis was conducted for each landscape pattern index. After selecting driving factors with p-values < 0.05, collinearity analysis was performed using OLS. Driving factors with a VIF greater than 7.5 were excluded. Finally, MGWR analysis was conducted at the district/county scale for the selected driving factors (Fig. 10).

Fig. 10
figure 10

MGWR results for CONNECT, PAFRAC, COHESION, AI, PLAND and PD (created using ArcGIS 10.8, https://www.esri.com; FRAGSTATS v4.2.1, https://www.fragstats.org/; MGWR 2.2, https://github.com/pysal/mgwr).

Full size image

For the CONNECT index, four regression factors were identified, with stronger explanatory power in the eastern region of the study area. NDVI, nighttime light coefficient, and GDP showed positive effects. The overall regression coefficient for NDVI was relatively high, with its significance gradually decreasing from the southwest to the east. The significance of the other two factors exhibited a pattern of higher influence in the east and lower influence in the west. The negative effect of NPP gradually weakened from the northwest to the southeast.

For the PAFRAC index, only two factors: slope and GDP—had regression values below 0.05. The regression coefficient for slope ranged from [0.68, 0.72], with high overall significance, especially in the central and eastern parts of the study area, where the widespread mountainous regions and significant slope variations resulted in irregular farmland shapes, leading to higher PAFRAC values. The negative effect of GDP gradually decreased from the north to the southeast.

The COHESION index is significantly influenced by five driving factors. Among these, urbanization rate exhibits a positive correlation, with its regression coefficient gradually decreasing from west to east. In contrast, nighttime light coefficient, distance from rural settlements, NDVI, and slope show negative correlations, with significance levels decreasing in that order. The absolute regression coefficients for nighttime light coefficient and distance from rural settlements decrease from the central and western regions to the east, with ranges of [−0.99, −0.67] and [−0.98, −0.18], respectively. In central and western Shandong, particularly in Jinan, a major economic hub, high nighttime light coefficient and sparse rural settlements result in larger values for the "distance from rural settlements" factor. Consequently, these areas exhibit higher fragmentation of cultivated land and weaker physical connectivity, leading to lower COHESION index values.

For the AI index, the explanatory power of its driving factors is strong, with an overall R2 of 0.91. Slope and nighttime light coefficient exert significant negative effects, with high absolute regression coefficients. In areas with steeper slopes, cultivated land patches are more fragmented, resulting in lower AI values. Similarly, regions with higher nighttime light coefficient, indicative of rapid economic development, also exhibit more fragmented cultivated land patches. In contrast, NPP and per capita arable land area positively influence the AI index. The effect of GDP varies spatially, showing a negative influence in the western regions of the study area and a positive influence in the eastern regions.

A MGWR analysis of the PLAND index yielded an R2 of 0.92, indicating strong explanatory power. Four driving factors exhibited negative correlations: GDP, slope, urbanization rate, and distance from rural settlements. Among these, GDP and slope had higher absolute regression coefficients, ranging from [−0.98, −0.12] and [−0.60, −0.51], respectively. In regions with high GDP, rapid economic development and urban expansion have encroached upon agricultural land, leading to a reduction in the PLAND index. Similarly, in mountainous areas with steeper slopes, forested land is more prevalent, and cultivated land is scarcer. Conversely, NPP, NDVI, DEM, and per capita arable land area were positively correlated with the PLAND index. As an indicator of net primary productivity, NPP is closely related to plant spatial distribution, with regions exhibiting higher NPP values generally having broader distributions of cultivated land. NDVI, as a vegetation index, also reflects the spatial distribution of cultivated land, leading to relatively higher regression coefficients for these two factors.

The regression analysis of the PD index demonstrated strong explanatory power for its driving factors, with a general pattern of lower explanatory strength in the east and higher strength in the west. Five factors—DEM, GDP, NPP, NDVI, and per capita arable land area—had negative impacts. Among these, DEM and GDP showed higher absolute regression coefficients, indicating that regions with higher elevation or GDP tend to have less cultivated land and lower density. In contrast, slope, nighttime light coefficient, and urbanization rate were positively correlated with the PD index. The slope regression coefficient ranged from [0.81, 0.94]. In areas with steep slopes, uneven terrain restricts the formation of contiguous cultivated land patches, resulting in higher fragmentation and density.

Discussion

Rationality of constructing the analysis model for cultivated land landscape fragmentation

This study proposes a model to assess cultivated land fragmentation in Shandong Province (1990–2023), integrating MSPA with landscape pattern indices. The integration of grid-based (fine-scale) and administrative (regional-scale) analyses enhances both the scientific rigor and methodological robustness of the results34. While MSPA is widely used in ecological network studies to identify ecological sources and classify core areas hierarchically62, such hierarchical analysis has rarely been applied in landscape fragmentation research. Unlike single-tier approaches, the stratified evaluation of core areas in this study offers clearer visualization of spatial disparities, temporal trends, and fragmentation intensity—supporting the rationale behind this methodological design. Regarding landscape pattern indices, class-level indices allow more detailed analysis of the distribution and fragmentation of specific land categories, while landscape-level indices are better suited for assessing overall structural patterns and heterogeneity21. Consequently, this study employs six class-level indices to systematically assess cultivated land attributes encompassing areal extent, density, connectivity, aggregation degree, and patch regularity. Landscape pattern indices are highly sensitive to boundary morphology and the delineation of the study area. Variations in grid scale dimensions and partition reference points may induce substantial discrepancies in index values for identical spatial extents. Therefore, district- and county-level administrative units were selected as analytical domains. This approach ensures methodological consistency through standardized boundaries and improves the practical relevance of results for land-use policy.

Policy recommendations derived from clustering analysis of cultivated land landscape patterns

In the realm of multi-indicator analysis, correlation testing has become an indispensable component, with most studies presenting analytical outcomes through numerical values or spatial relationships36,38. Recent methodological advancements have witnessed increasing scholarly efforts to integrate these dual analytical dimensions40. This investigation quantifies interrelationships among landscape pattern indices through Pearson correlation analysis during cultivated land fragmentation assessment, while simultaneously visualizing spatial interdependencies via bivariate spatial autocorrelation analysis. The consistent trends across both methods highlight their complementary strengths and provide a more comprehensive analysis. Most current studies on landscape pattern relationships stop at this stage34. To advance beyond conventional limitations, this study explores deeper structural connections within landscape patterns. Although clustering techniques are widely used in ecosystem service studies78, they are rarely applied in fragmentation research. This study applies the Self-Organizing Map (SOM) method to bridge that gap. The resultant cluster typology is visualized through sector diagrams, effectively revealing latent associations among cultivated landscape pattern indices. This methodological innovation facilitates more nuanced policy recommendations grounded in multidimensional pattern recognition. Clustering analysis of landscape pattern indices provides a basis for corresponding policy recommendations. The six clusters can be divided into three groups:

Group 1 (Clusters B and E) is characterized by small proportions of cultivated land and irregular patch shapes, making these areas unsuitable for agricultural production. Geographically, these areas are located in Jinan (provincial capital) and Qingdao (major port city). Future policies should focus on concentrated economic development. If demand for construction land rises, cultivated land in these areas may be repurposed for development, with compensation achieved through land reclamation in more agriculturally suitable regions.

Group 2 (Cluster A), primarily distributed in central and eastern coastal Shandong, exhibits moderate cultivated land proportions with good connectivity. These areas should adopt coordinated development of manufacturing and agriculture. Recommended policies include improving agricultural output through high-standard farmland development, land consolidation, soil-specific improvements, and comprehensive saline-alkali remediation, particularly in the Yellow River Delta.

Group 3 (Clusters C, D, and F), located in western, northwestern, and central-eastern Shandong, features high cultivated land proportions, aggregated patches, and strong connectivity, suitable for large-scale mechanized farming. Recommendations include expanding high-standard farmland with complete drainage and hardened field roads, boosting yields of food and cash crops, and optimizing agro-processing supply chains that integrate production, processing, and marketing.

The longitudinal analysis of clustering regime spatial distribution patterns from 1990 to 2023 reveals distinct evolutionary trajectories, underscoring the critical importance of extended temporal investigations79. Sankey diagrams of transition matrices for the six cluster types show an intensifying fragmentation trend in Shandong’s cultivated landscapes over time (Fig. 9). Notably, recent governmental cropland protection initiatives exhibit initial effectiveness in moderating fragmentation transition rates (High-standard Farmland Construction Policy, Integrated Whole-region Land Consolidation Policy, Three Zones and Three Lines Delineation Policy, and related policies). This comprehensive assessment establishes that conventional correlation analysis, when augmented with clustering techniques and transition matrix examination, enables deeper interrogation of intrinsic landscape pattern relationships40. The multidimensional analytical framework successfully disentangles complex fragmentation dynamics while maintaining methodological parsimony.

Regression analysis of driving factors for cultivated landscape fragmentation

Initial morphological and landscape pattern analyses revealed persistently intensifying fragmentation of cultivated land in Shandong Province, a condition that significantly compromises ecosystem service provision and agricultural productivity12. Hence, systematic examination of driving factors becomes imperative to elucidate the intrinsic mechanisms governing landscape ecological evolution in cultivated areas80. Following established research protocols and expert consultations, we initially selected 11 accessible explanatory variables for MGWR analysis specific to Shandong Province. Notably, road density and crop yield demonstrated no significant correlations with any of the six landscape pattern indices through MGWR examination. Subsequent diagnostic analysis identified that the minimal inter-district/county variation in road network density—resulting from incorporating multi-level road data—likely contributed to this statistical insignificance. Similarly, limited regional disparity in crop yields across the province produced elevated p-values. Future investigations should consider implementing hierarchical road data screening protocols. Critical determinants including slope gradient, proximity to rural settlements, and GDP exhibited strong explanatory power in regression models, consistent with previous empirical findings81,82.

At the outset of the MGWR analysis, explanatory variables with p-values less than 0.05 were initially selected to ensure statistical significance. Multicollinearity among these variables was then assessed using OLS, and those with VIF exceeding 7.5 were excluded. This screening process ensured that the final set of variables used in the regression provided a reliable overall explanation of the landscape pattern indices. However, R2 values derived from the MGWR models for six landscape metrics revealed substantial variation. While AI [0.85,0.97], PLAND [0.92,0.96], and PD [0.82,0.94] exhibited consistently high explanatory power, CONNECT [0.43,0.97], PAFRAC [0.48,0.62], and COHESION [0.30,0.93] showed markedly lower local R2 values in certain areas. These spatial discrepancies suggest that, in some regions, the selected driving factors had limited ability to explain the variability in these particular indices. Such heterogeneity is to be expected in studies covering large and diverse geographic areas, where the influence of driving forces naturally varies across space. Consequently, localized inconsistencies in MGWR model performance reflect the inherent spatial complexity of landscape patterns. This spatial heterogeneity in model performance is expected in a study area of considerable geographic extent, where the influence of driving factors is unlikely to be uniform. As such, it is not uncommon for MGWR results to reveal subregions where the local regressions offer limited explanatory power, reflecting the underlying spatial complexity of landscape dynamics.

Two approaches may be considered to address this issue. First, future research could adopt a finer spatial scale of analysis. Due to constraints in the availability of driving factor data, the present study was conducted at the county level. However, counties are relatively large spatial units, which limits the ability to capture internal heterogeneity. For raster-based variables, the analysis required calculating the average value of all grid cells within each county using ArcGIS, a process that inevitably smooths over spatial variation within the unit. To address this, future research could use smaller administrative units such as towns or sub-districts, which would not only better reflect spatial heterogeneity but also increase the number of regression samples, thereby enhancing the robustness of the analysis. Where data availability remains a constraint, field surveys could be conducted to obtain localized information on driving factors. Second, regions exhibiting low R2 values in the MGWR analysis warrant targeted field investigation. On-the-ground observations could help identify context-specific influences on cropland distribution, enabling the refinement of existing variables and the introduction of additional factors that may have been overlooked in the current study.

Limitations and future research directions

This study is subject to several limitations that warrant further attention in future research. First, the vector data on cropland distribution are classified government information and thus not publicly accessible. Due to these limitations, this study employs raster data interpreted from remote sensing imagery by Wuhan University. While this dataset provides a valuable alternative, it inevitably contains some discrepancies compared to actual cropland distributions, which may introduce minor biases into the analysis. For instance, in certain regions of Shandong Province, greenhouses or plastic tunnels are temporarily constructed during specific crop growth stages to retain heat and are dismantled once temperatures rise. These structures can easily be misclassified as built-up land during remote sensing interpretation, leading to deviations between interpreted data and official vector datasets. Such discrepancies may affect the calculated values of cropland landscape pattern indices to some extent. However, given the large spatial scale and extended temporal scope of this study, the emphasis lies on capturing overarching trends in spatial pattern evolution rather than fine-scale detail. While the influence of data precision is acknowledged, it is not considered severe in this context. Nevertheless, to enhance scientific rigor, future studies should seek to incorporate higher-precision datasets wherever possible.

Second, the exclusive focus on landscape configuration metrics overlooks critical quality dimensions of cultivated land. Recent policy implementations have substantially enhanced soil productivity across regions, necessitating future syntheses of landscape pattern analysis with cultivated land quality assessments. Such multidimensional approaches incorporating improved data sources and analytical dimensions could comprehensively characterize Shandong’s agricultural land status.

Conclusions

  1. 1.

    From 1990 to 2023, Shandong Province exhibited progressive reductions in core cultivated areas, pore spaces, and island-shaped patches. The proportion of primary core areas decreased substantially from 82.44 to 30.28%, while other morphological landscape types showed cumulative expansion. These changes collectively indicate diminished landscape connectivity and intensified fragmentation throughout the study period.

  2. 2.

    Multivariate analysis of six landscape pattern indices confirms escalating fragmentation of cultivated landscapes from 1990 to 2023. Central and southeastern coastal regions demonstrated lower cultivated area ratios, elevated patch densities, reduced connectivity indices, diminished aggregation metrics, and heightened shape irregularity. Conversely, western regions maintained superior agricultural continuity with higher cultivated area ratios values, concentrated patch distributions, and enhanced aggregation patterns.

  3. 3.

    Bivariate spatial autocorrelation and Pearson correlation analyses yielded congruent trend patterns. Pearson coefficients revealed the strongest negative correlation between PLAND and PD (mean coefficient = −0.9325 across four periods), while PLAND-AI showed the highest positive correlation (mean coefficient = 0.9225). Cluster analysis identified Clusters C, F, and D as optimal for cultivation, characterized by large contiguous areas, high aggregation, and superior connectivity. During the study period, Cluster F expanded from 17.6 to 27.9% of total cultivated land, whereas Clusters C and D exhibited proportional declines.

  4. 4.

    MGWR identifies key driver relationships: CONNECT-NDVI (positive correlation); PAFRAC and PD-slope (positive correlation); COHESION-nighttime light coefficient (negative correlation); AI-slope (negative correlation); and PLAND-GDP (negative correlation).