Analysis of farmland fragmentation in China Modernization Demonstration Zone since “Reform and Openness”: a case study of South Jiangsu Province

Farmland is a fundamental resource for human survival and development. However, farmland fragmentation has become a serious problem, causing ecological damage and low crop production efficiency in many parts of the world. Based on remote sensing and socioeconomic data, we used landscape pattern indices, Morphological Spatial Pattern Analysis (MSPA), and Markov chain models to analyze the temporal and spatial pattern changes in farmland in South Jiangsu Province (the first “Modernization Demonstration Zone” in China) during 1985–2010. Our results demonstrated that the total farmland area decreased by ca. 24% and the farmland pattern became fragmented during 1985–2008: core farmland decreased and islet farmland increased. Additionally, the farmland patch density (PD) increased and three other landscape indices (NLSI, MESH, and COHESION) showed significant decreases. Although the fragmentation rate slowed after 2008, the convergence rate to a stationary farmland distribution became faster, and transitions tended to be less deterministic after 2000. Economic and population growth and policy changes positively contributed to this phenomenon. Therefore, the primary task of farmland protection should be to preserve contiguous farmlands and reduce scattered patches in order to promote farmland connectivity.


Results and Discussions
Farmland changes. Our classification results (Table 1) demonstrate that the total farmland area in South Jiangsu Province decreased from 17915 km 2 to 13638.64 km 2 during 1985-2010. The net reduction is about 4300 km 2 and the total proportion of farmland decreased from 64.56% to 49.10%. The average annual drop in farmland area increased from 95.56 km 2 (1985-1995) to 389.96 km 2 (2005-2008). Similarly, the average annual decrement rate increased from 0.53% (1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995) Table 2 for details) in this area show the progress of farmland fragmentation (Fig. 1c,d). During 1985-2010, the proportion of core farmland (representing the main stable and contiguous farmland) decreased from 65.8% to 54.37%. The boundary (perforation and edge classes) farmland increased from 29.54% to 37.02%, which implies that the core farmland acquired a more irregular shape. Overall, the relative proportion of islet farmland (interspersed farmland unsuitable for mass production) almost doubled, which directly reflected the increased fragmentation. Although the proportion of connectors (loop and bridge classes) increased from 2.4% to 4.2%, demonstrating the improved connectivity between core farmlands, the simultaneous increment (from 2.11% to 4.19%) of branch farmland (randomly attached to boundary farmland) indicated that the fragmentation was still increasing. In addition to the results presented above, MSPA allows for direct observation of farmland fragmentation, as shown in Fig. 2, which clearly shows that farmland area has sharply deceased, and become Figure 1. Landscape index changes (a,b; top) for farmland, including the patch density (PD), the normalized landscape shape index (NLSI), COHESION, and the effective MESH size (MESH). Changes in the proportion of different MSPA types for the entire farmland area (c,d; middle). Changes in the number of practitioners (P), agricultural practitioners (AP), the relative proportion of AP (PAP), the regional Gross Domestic Production (GDP), the regional Gross Agricultural Production (GAP) and the proportion of GAP (PGAP) (e,f; bottom) in South Jiangsu Province during 1985-2010. The figure was generated by L.Z. and Y.P. using Origin 8.5 (http://www.originlab.com/). irregular and fragmented. The majority of fragmented farmland in the study area ( Fig. 2) were in the southeast, close to Shanghai (the economic and financial center of China), and in the northwest near Nanjing (the capital of Jiangsu Province). These two cities affect the surrounding regions significantly and directly influence the farmland configuration in South Jiangsu Province.
Markov chain models for MSPA classes. By applying a Markov chain model to the MSPA results, we can see the conversions between different morphological classes before and after 2000 from the transition probability matrices (Supplementary Table S1, S2). We can also map the transition process of core farmland in a small region (Fig. 3a), and the transition areas where core farmland was converted to other classes across the entire South Jiangsu (Fig. 3b). During 1985-2010, transitions mainly occurred within the same class, as shown by bigger diagonal values (> 0.5) in the transition probability matrix. Core farmland was the most stable part, with about 80% remaining unchanged, which shows the stability of large continuous farmlands. Transitions between different classes (but within the same type) mainly occurred from the perforation to edge class (boundary type), and from the loop to bridge class (connector type), as shown by the larger off-diagonal values: transition probabilities were 0.1816 and 0.1861, respectively, before 2000, and 0.2690 and 0.1791 after 2000. These two transitions clearly show the farmland fragmentation from the sides (Supplementary Figure S1). Almost 75% of transitions to non-farmland occurred from islet and branch farmland, but transitions from non-farmland to farmland were extremely rare. Apart from branch and islet farmland, almost every farmland class appeared more stable before 2000. These results combine to indicate a more fragmented and poorly connected scenario after 2000.
The convergence rate (ρ) and normalized entropy (H(P)) of the transition matrix during 1985-2000 was 1.1647 and 0.4093, respectively. For the next decade (2000-2010), ρ was 1.2256, which indicates a faster convergence rate to a stationary farmland distribution, and more frequent transitions between different classes. H(P) for this decade was 0.3848, which indicates less deterministic and more unpredictable Markov transitions compared with the previous period. H(P) for 1985-2010 was approximately 0.39, indicating a partially deterministic transition process. The larger convergence rate and smaller entropy after 2000 indicate that the farmland fragmentation rate increased and was not well controlled.

Driving forces behind farmland loss and fragmentation.
Although correlations between the chosen factors (see Methods), islet farmlands and the PD were not significant, significant correlations were seen between factors, and other classes and indices, indicating the appropriateness of our chosen factors ( Table 3). The regional GDP grew from 43.3 billion Yuan (the monetary unit of China) to 2506.74 billion Yuan during 1985-2010, while the regional GAP grew from 6.24 billion Yuan to just 58.43 billion Yuan (Supplementary Table S3, Fig. 1e,f). The proportion of GAP decreased from 14.43% to 2.33%, accompanied by a significant loss of local agricultural practitioners caused by the regional South Jiangsu Model 36 (a kind of economic development model which relied mainly on a collective economy). The economic barycenter transferred to the secondary and tertiary industries, causing significant loss and fragmentation of farmland.

Type Explanation Class Definition
Core "Matrix" (main, stable and contiguous) farmland Core Foreground pixels whose distance to the background is greater than the given edge width D  Our results also showed different farmland loss and fragmentation rates during 1985-2010 (Figs 1,2). The rate was slow during 1985-1995, sharply increased after 1995, but slowed down after 2008. Combined with related policies, we found that China had been restoring the national economy from the beginning of the "Reform and Openness" program up to the late 1980s. Since 1992, China has established a socialist market economic system, and has begun to develop the economy energetically. South Jiangsu Province has attracted investment and developed township enterprise as a result of grants from the "Reform and Openness"; however, as the economy developed and urbanization accelerated5, farmland was rapidly lost and became fragmented. To tackle this issue, several policies were implemented to protect farmland,  Suggestions for farmland protection. Avoiding fragmentation is a key aspect of farmland preservation 12 . China should reinforce the implementation of the "Agricultural land consolidation" program, including the merging of fragmented farmland patches. Additionally, a balance must be kept between farmland occupation and compensation, and the protection of basic farmland must also be implemented. From the perspective of landscape patterns, the protection of core farmland should be the first priority, alongside the reduction of boundary and branch farmland to keep the farmland shapes regular. Small islet farmland should be reduced or incorporated into the core farmland. Networks and the "corridors" of landscape are also essential for increasing the landscape's connectivity 37 , implying that an increase in connector farmland could strengthen the COHESION of farmland landscape. With the current vigorous economic development, policy makers should not only protect high-quality and contiguous farmlands, but should also continue to promote the agricultural development, raise the level of agricultural modernization 34 , and adjust the relevant land use policies according to the local situation. Ideally, both the total amount and the morphological structure of farmland should remain balanced and stationary.

Conclusions
This study analyzed farmland loss and farmland pattern changes using traditional landscape indices in South Jiangsu Province during 1985-2010. The increase in the PD and the decrease in the NLSI, and the MESH and COHESION indices demonstrated considerable farmland fragmentation. We also applied a new MSPA method that can visually present farmland patterns changes and guarantee the accuracy of farmland landscape analysis at the pixel level. Direct images showed a decrease in core farmland and an increase in other types of farmland, revealing that the most fragmented areas were in the northwest and the southeast of the study area. Overall, farmland in this area diminished and fragmented faster before 2008. By integrating a Markov chain model and the MSPA results, we could clearly observe the transition processes for the different farmland classes and how these transitions influenced (or aggravated) farmland fragmentation. We observed a relatively stable farmland configuration but obvious transitions to non-farmland. Results after 2000 indicated a faster and less deterministic progression towards a stationary farmland distribution. This situation had direct and significant relationships with socioeconomic, population, and policy factors, such as the regional GDP and the number of agricultural practitioners.
Owing to the ecological and economic damage caused by fragmentation, high-quality and contiguous farmlands should be protected. From a landscape ecology perspective, scattered islet, boundary, and branch farmland should be reduced to protect the farmland shape and increase the connectivity of farmland in the study area. Future studies should focus on farmland morphological changes after 2010, when South Jiangsu Province expedited economic restructuring that does not rely predominantly on industries that occupied large areas of farmland. Additionally, support vector machine (SVM) classification system could be employed in future studies to improve classification accuracy and reduce the amount of labor involved in classification.

Methods
Study area. South Jiangsu Province (118.35°E-121.33°E, 30.76°N-32.61°N) is located in the southeast coastal region of China, spanning 5 cities (Nanjing, Zhenjiang, Changzhou, Wuxi, and Suzhou) and covering an area of 27,872 km 2 (terrain: 50.45% plains, 28.4% hills, and 21.12% water). The region belongs to the East Asian monsoon climate zone, which is characterized by cold winters, hot summers, and a generally humid environment. The mean precipitation is 1000-1280 mm/yr, and in the rainy season (June and July), it is around 250 mm. The study area has rich water resources, including Lake Taihu (one of China's five largest freshwater lakes), and the Yangtze River, which traverses the territory from east to west and is 432 km long (Fig. 4). The superior geographical location and climate of this region combine to provide an excellent environment for agricultural production.
The resident population of the region was 32.56 million in 2011, and the population density was 1166 people/km 2 , 8.3 times the national average. South Jiangsu has developed a unique regional economic development model: the South Jiangsu Model. This model is characterized by township and village enterprises, and is regarded as one of the most successful modes of rural territorial development in China. Therefore, the regional GDP reached 3.64 × 10 12 Yuan, 6.4% of China's total GDP. Per capita GDP reached 1.14 × 10 5 Yuan in 2013, ranking 41 st in the world according to the International Monetary Fund (IMF). The region has seen some of the fastest growth in its economy and traffic network in China since the "Reform and Openness", and its urbanization rate reached 70.8% in 2013. The South Jiangsu Modernization Demonstration Zone, the first (currently the only) regional zone taking modernization as its theme, was formally established in 2013.
South Jiangsu Province contained abundant farmland resources and the crop sowing area was 1.24 × 10 6 ha in 2011. Although it is one of the main grain-producing areas in China, farmland area and food production (mainly rice with some wheat) have continued to decrease since the 1990s, resulting in a decrease in grain possession per capita of ca. 30%.
Remote sensing data and data preprocessing. The Table S4). The starting point for this study was the initial period of the "Reform and Openness" program in the 1980s (1985, also the final year of the 6 th Five-Year Economic Plan) and the selected study years marked the ends of the 7-10 th Five-Year Economic Plan periods, allowing an assessment of the economic effects on farmland. The year 2008 was selected because (1) the economy developed faster during this year, which may have aggravated farmland morphological changes, and (2) the concept of permanent basic farmland was first introduced.
For each image, we performed image preprocessing using ENVI 4.7, which involves geometric correction, radiometric calibration, atmospheric correction, and image mosaicking and clipping. A few clouds were excluded by the Haze Tool (an IDL plug-in in ENVI). Then, we processed a maximum-likelihood supervised classification and included six main land use types: farmland, built-up land, forest, grassland, water bodies, and other land (Supplementary Figure S2). We assessed the classification accuracy using 300 random small ROIs (region of interest) selected from the land use data from Data Sharing Infrastructure of Earth System Science (http://www.geodata.cn/) and Google Maps. Our results from confusion matrices showed that the total accuracy of each classified image was around 87%, which was better than the recommended 85% 41 , and the kappa coefficient was around 0.7 (Table 4)  The main factors behind farmland changes may have been socioeconomic, cultural, and political trends 21,42 . We therefore obtained socioeconomic data from the Statistical Yearbook of each city (e.g. Nanjing Statistics Bureau website: http://www.njtj.gov.cn/) to study the driving forces behind farmland pattern changes during 1985-2010. The relevant factors chosen for this study were: (1) regional Gross Domestic Production (GDP), (2) regional Gross Agricultural Production (GAP), (3) proportion of GAP (PGAP), (4) total number of Practitioners (P), (5) Agricultural practitioners (AP), and (6) proportion of AP (PAP). A bivariate Pearson correlation analysis was used to investigate correlations between these factors, and landscape indices and morphological classes using SPSS 21.0.

Landscape pattern indices.
Landscape pattern indices can illustrate the relationship between landscape patterns and ecological processes at the landscape level 43 , and advance our understanding of the ecological functioning of landscapes 44 . Since the selection of specific indices for different landscapes appears idiosyncratic, and no individual index can capture the full complexity of spatial patterns, highly contingent and multiple sets of indices are required 45 . In this study, we selected the following four indices to reveal the morphological changes of farmland comprehensively: (1) the patch density (PD), (2) the normalized landscape shape index (NLSI), (3) the COHESION index (COHESION), and (4) the effective MESH size (MESH) These indices are calculated by FRAGSTATS 46 , and are defined and explained in Supplementary Table S5.
Morphological spatial pattern analysis (MSPA). MSPA is based on mathematical morphology and relies on two fundamental morphological operations: erosion and dilation 47 . MSPA can automatically segment preprocessed binary images at the pixel level into different types or classes representing different sizes, shapes, and connectivity 22 . This allows direct observation of the fragmentation process and the relationships between different classes, aiding the interpretation of results. MSPA uses five mutually exclusive types (or seven classes) for foreground pixels: core, islet, connector (including the bridge and loop classes), boundary (including the perforation and edge classes), and branch 22 . These types are defined in Table 2 and illustrated in Supplementary Figure S3.
MSPA is a highly scale-dependent approach, and so the relative abundance of MSPA classes is related to the density of the focal class (F), the level of contagion (H), the edge width of the morphological windows (D), and the connectivity rule (R) 48 . For each given map, F and H are set, so only changes in D and R will affect the MSPA results. R can be set to four neighbors (the pixels' edges are directly connected to a central pixel) or eight neighbors (the pixels' edges or corners are connected to a central pixel). The edge width (D), defined as a b 2 2 + (a and b are the distance along the horizontal and vertical axes respectively), is used to determine the different classes in MSPA. R and D determine the quantity of each class, but only R decides the general trends of the relative proportions and shapes of different classes 49 . A smaller value of D can provide more differentiation among structural classes and more information on landscape patterns 50 .
During 1985-2010, the proportions of farmland area (F) in South Jiangsu ranged from 49% to 65%, and the contagion index (H) was around 47-50. According to an experimental analysis of random maps with contagion and of real maps 48 , we chose a connectivity rule (R) set to 8 neighbors, and set D = 4 (i.e., edge width = 120 m) for this study. According to the characteristics of different classes, a decrease in the amount of core and connector farmland, and an increase of other types can reveal the fragmentation directly.
Markov chain models. Markov chain models can analyze temporal landscape changes 51 , and have been used previously in landscape ecology studies to identify the direction and rate of the transition between patches, corridors, and matrices 52,53 . In this study, we used discrete-time, finite-state, homogeneous (stationary) Markov chain models to analyze the succession of MSPA classes in 1985-2000 and 2000-2010. Our model viewed the landscape as a large set of pixels, and the state of each pixel  was identified by the occupying land use type (Supplementary Panel S1). After obtaining the transition probability matrix P, we used two variables 52 to analyze the matrices: the convergence rate ρ and the entropy H(P). A smaller ρ indicates a slower convergence rate to a stationary distribution. If H(P) is closer to 0, this indicates a more deterministic Markov transition, and values closer to 1 indicate a more random Markov transition.