A non-grain production on cropland spatiotemporal change detection method based on Landsat time-series data
基于 Landsat 时间序列数据的农田非粮生产时空变化检测方法
首次发布：2024 年 4 月 3 日 https://doi.org/10.1002/ldr.5113
Tingting He, Suqin Jiang, and Wu Xiao contributed equally to this study and should be considered co-first authors.
Tingting He、Suqin Jiang 和 Wu Xiao 对本研究做出了同等贡献，应被视为共同第一作者。
Abstract 摘要
Global food security is being threatened by the reduction of high-quality cropland, extreme weather events, and the uncertainty of food supply chains. The globalization of agricultural trade has elevated the diversification of non-grain production (NGP) on cultivated land to a prominent strategy for poverty alleviation in numerous developing nations. Its rapid expansion has engendered a multitude of deleterious consequences on both food security and ecological stability. NGP in China is becoming very common in the process of rapid urbanization, threatening national food security. To better understand the causal mechanisms and enable governments to balance food security and rural development, it is crucial to have a clear understanding of the spatiotemporal dynamics of NGP using remote sensing. Yet knowledge gaps remain concerning how to use remote sensing to track human-dominated or -induced long-term cultivated land changes. Our study proposed a method for detecting the spatiotemporal evolution of NGP based on Landsat time-series data under the Google Earth Engine platform. This approach was proposed by (1) obtaining the union of cultivated lands from multiple landcover products to minimize the cultivated land omission, (2) constructing multi-index dynamic trend rules for 3 representative types of NGP and obtaining results at the pixel level, while adopting the continuous change detection and classification algorithm to Landsat time series (1986–2022) to determine when the most recent change occurred, (3) minimizing the noise by object-oriented land use–land cover classification and mode filter approaches, and (4) mapping the spatiotemporal distribution of NGP. The proposed methodology was tested in Jiashan, located in Zhejiang Province (eastern China), where NGP is widespread. We achieved a high overall accuracy of 95.67% for NGP type detection and an overall accuracy of 85.26% for change detection of time. The results indicated a continued increasing pattern of NGP in Jiashan from 1986 to 2022, with the cumulative percentage of NGP increasing from 0.02% to 20.69%. This study highlights the utilization of time-series data to document essential NGP information for evaluating food security in China and the method is well-suited for large-scale mapping due to its automatic manner.
全球粮食安全正受到优质耕地减少、极端天气事件和粮食供应链不确定性的威胁。农业贸易的全球化已将耕地非谷物生产的多样化（NGP）提升为许多发展中国家减贫的一项重要战略。它的迅速扩张给粮食安全和生态稳定带来了许多有害后果。中国 NGP 在快速城市化过程中变得非常普遍，威胁着国家粮食安全。为了更好地了解因果机制并使政府能够平衡粮食安全和农村发展，利用遥感清楚地了解 NGP 的时空动态至关重要。然而，在如何利用遥感跟踪人类主导或引起的长期耕地变化方面，仍然存在知识差距。我们的研究提出了一种在 Google Earth 引擎平台下基于 Landsat 时间序列数据检测 NGP 时空演变的方法。该方法是通过（1）从多个土地覆盖产品中获得耕地的联合，以最小化耕地遗漏，（2）为 3 种代表性类型的 NGP 构建多指标动态趋势规则并获得像素级结果，同时对 Landsat 时间序列（1986-2022 年）采用连续变化检测和分类算法来确定最近变化发生的时间，（3）通过面向对象的土地利用-土地覆盖分类和模式滤波方法来最小化噪声，以及（4）绘制 NGP 的时空分布。 所提出的方法在位于浙江省（中国东部）的嘉善进行了测试，那里 NGP 很普遍。我们实现了 NGP 类型检测的 95.67%的高总体准确率和时间变化检测的 85.26%的总体准确率。结果表明，从 1986 年到 2022 年，嘉善 NGP 呈持续增加的模式，NGP 的累计百分比从 0.02%增加到 20.69%。这项研究强调了利用时间序列数据来记录评估中国粮食安全的基本 NGP 信息，并且由于其自动方式，该方法非常适合大规模制图。
1 INTRODUCTION 1 导言
According to the “2023 World Food Security and Nutrition Report,” between 691 and 783 million people worldwide experienced hunger in 2022, with over 90 million people being affected by long-term hunger (FAO et al., 2023). Conflicts, extreme weather events, diseases, and economic shocks are exacerbating the main driving factors of food insecurity, and stable food production is crucial for global food security and sustainable development (Liang et al., 2023). There are two types of production methods on arable land: grain production and non-grain production (NGP) (Su et al., 2020; Wang & Dai, 2022). In cases where cultivated land is utilized for NGP, agricultural behaviors still take place, but staple food crops are not cultivated. In nations and regions grappling with a combination of water scarcity and intense agricultural land use conflicts, the rapid expansion of cash crops is encroaching on a substantial portion of cropland traditionally used for staple food production (Caldarelli & Gilio, 2018; Chiarelli et al., 2018; Defante et al., 2018; Li et al., 2018; Skiba et al., 2020). This trend presents a significant challenge to achieving local self-sufficiency in staple food (Huang, 2022; Lim, 2023; Qianru et al., 2023).
根据《2023 年世界粮食安全和营养报告》，2022 年全球有 6.91 至 7.83 亿人经历饥饿，其中超过 9000 万人受到长期饥饿的影响（粮农组织等人，2023）。冲突、极端天气事件、疾病和经济冲击正在加剧粮食不安全的主要驱动因素，稳定的粮食生产对全球粮食安全和可持续发展至关重要（Liang et al.，2023）。耕地上有两种生产方式：粮食生产和非粮食生产（NGP）（苏等，2020；王&戴，2022）。在将耕地用于 NGP 的情况下，农业行为仍然发生，但不种植主食作物。在努力应对水资源短缺和激烈的农业土地使用冲突的国家和地区，经济作物的快速扩张正在侵占传统上用于主食生产的很大一部分农田（Caldarelli&Gilio，2018；Chiarelli 等人，2018；Defante 等人，2018；李等，2018；Skiba 等人，2020）。这一趋势对实现当地主食自给自足提出了重大挑战（黄，2022；林，2023；倩茹等人，2023）。
The large domestic food demand and limited cropland resources in China determine the extremely prominent importance of cultivated land protection and food security for economic development and social stability (Su et al., 2019; Ziegler et al., 2009). The loss of cultivated land productivity caused by rapid urbanization has raised higher demands for cultivated land protection. In China, over the past decade, the phenomenon of NGP driven by comparative advantages has become increasingly serious. The hidden loss of grain productivity caused by NGP has far exceeded the explicit loss caused by non-agriculturalization (Ren et al., 2023; Wang & Dai, 2022). In addition, NGP can also lead to issues such as soil erosion (El Kateb et al., 2013; Hu et al., 2017) and non-point source pollution (Chatvijitkul et al., 2017), posing challenges to ecological security (Foley et al., 2005). To ensure an effective staple food supply and rice self-sufficiency, it is necessary to prioritize the limited cultivated land resources for food production to safeguard an adequate provision of staple food (Chen et al., 2021).
中国国内巨大的粮食需求和有限的耕地资源决定了耕地保护和粮食安全对经济发展和社会稳定的重要性极其突出（苏等，2019；齐格勒等人，2009 年）。快速城市化带来的耕地生产力损失，对耕地保护提出了更高的要求。在中国，近十年来，比较优势驱动下的 NGP 现象日益严重。NGP 造成的粮食生产力隐性损失已经远远超过非农化造成的显性损失（任等，2023；王戴，2022）。此外，NGP 还可能导致水土流失等问题（El Kateb 等人，2013 年；胡等人，2017 年）和面源污染（Chatvijitkul 等人，2017 年），对生态安全构成挑战（Foley 等人，2005 年）。为了确保有效的主粮供应和水稻自给自足，有必要将有限的耕地资源优先用于粮食生产，以保障主粮的充足供应（Chen et al.，2021）。
The identification of NGP information from cultivated land serves as the fundamental basis for other researches. Enhancing the accuracy and timeliness of NGP information is crucial in providing robust data for subsequent analysis and governance. Currently, there are two main approaches to identifying NGP. The first approach involves questionnaire surveys or statistical data to quantify NGP information, while the second approach utilizes remote sensing techniques for NGP detection. Most studies quantitatively measure NGP using statistical data collected by provincial, municipal, and county-level statistical agencies. Some studies have used indicators such as “non-grain sown area to crop sown area” (Guo & Wang, 2021; Li, 2017; Ren et al., 2023; Yu et al., 2015; Zhang et al., 2022) and “the ratio of the grain-sowing area to cultivated land and the multiple cropping index” (Sun et al., 2021; Wang & Dai, 2022) to describe the rate of NGP. These methods have gaps in temporal coverage, and cannot fully capture the spatial characteristics of NGP. The method that directly extracts NGP information from China's 2020 Land Survey Database has been shown to effectively display the spatial distribution of NGP in the study area up to the time of China's 2020 Land Survey (Guan et al., 2021; Liang et al., 2023; Yang & Zhang, 2021). However, it should be noted that the confidentiality of the data limits its application in research.
耕地 NGP 信息的识别是其他研究的基础。提高 NGP 信息的准确性和及时性对于为后续分析和治理提供可靠的数据至关重要。目前，有两种主要的方法来识别 NGP。第一种方法涉及问卷调查或统计数据来量化 NGP 信息，而第二种方法利用遥感技术来检测 NGP。大多数研究使用省、市和县级统计机构收集的统计数据对 NGP 进行定量测量。一些研究使用了“非粮播种面积与作物播种面积之比”等指标（郭&王，2021；李，2017；任等，2023；于等，2015；张等，2022）和“粮食播种面积与耕地之比和复种指数”（孙等，2021；王&戴，2022）来描述 NGP 的比率。这些方法在时间覆盖上存在差距，不能完全捕捉 NGP 的空间特征。从中国 2020 年土地调查数据库中直接提取 NGP 信息的方法已被证明可以有效地显示截至中国 2020 年土地调查时 NGP 在研究区域的空间分布（关等，2021；梁等，2023；杨&张，2021）。然而，应该注意的是，数据的保密性限制了其在研究中的应用。
Remote sensing has significant advantages in acquiring information on large-scale land cover changes, allowing for better spatiotemporal evolution of land cover, and it has gradually become the mainstream method for obtaining land cover information (Buchner et al., 2020; Chai & Li, 2023). Some researchers have adopted methods based on landcover classification to extract NGP information (Su et al., 2019; Xiao et al., 2015; Yang & Zhang, 2021). However, this approach may face challenges in ensuring temporal consistency due to issues such as image noise and classification algorithms, which can affect the accuracy of classification results in each period and may magnify uncertainties in land cover change trajectories, making it difficult to determine whether the observed changes are real or result from misclassification in a particular image (Xu et al., 2023). Some other scholars have used decision tree models (Zhang et al., 2023) or existing datasets of food crop planting (Zhu et al., 2022), which are all limited in their ability to distinguish specific types of NGP.
遥感在获取大范围土地覆盖变化信息方面具有显著优势，可以更好地进行土地覆盖时空演变，逐渐成为获取土地覆盖信息的主流方法（Buchner et al.，2020；柴李，2023）。一些研究人员采用了基于土地覆盖分类的方法来提取 NGP 信息（苏等，2019；肖等，2015；杨&张，2021）。然而，由于图像噪声和分类算法等问题，这种方法在确保时间一致性方面可能面临挑战，这可能会影响每个时期分类结果的准确性，并可能放大土地覆盖变化轨迹的不确定性，从而难以确定观察到的变化是真实的还是由特定图像中的错误分类引起的（Xu 等人，2023）。其他一些学者使用了决策树模型（张等，2023）或现有的粮食作物种植数据集（朱等，2022），这些模型在区分特定类型 NGP 的能力方面都受到限制。
None of the above-mentioned methods can accurately provide the spatiotemporal dynamics information for different categories of NGP, it is necessary to propose new perspectives and methods. Time series change detection method can capture the characteristics of landcover changes in spatiotemporal scope (Brown et al., 2020; Liu et al., 2019; Sulla-Menashe et al., 2016; Vogelmann et al., 2016), which typically employs single-band quantitative parameters, such as Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and so on, as input data rather than multispectral images. Thanks to the fact that the acquisition cost of historical remote sensing images has been greatly reduced, the time-series images with simultaneous phases (e.g., monthly, or seasonal time-series images over multiple years) are more conveniently used to effectively explore the temporal dynamics of land features. Currently, there are two main approaches for time series change detection. The first approach involves using annual images to generate a time series spanning multiple years. Typical algorithms for this approach include the Vegetation Change Tracker algorithm (Huang et al., 2010) and the LandTrendr algorithm for disturbance and recovery trend detection (Kennedy et al., 2010). The second approach first uses all available images each year to form an intra-year time series, and then concatenates it into a multi-year time series, and is represented by algorithms such as the Breaks For Additive Seasonal and Trend algorithm (Verbesselt et al., 2010) and the Continuous Change Detection and Classification algorithm (CCDC) (Zhu & Woodcock, 2014). Currently, only a few researchers have applied temporal remote sensing methods for identifying NGP conversion information in cultivated land (Yu, 2020).
上述方法都不能准确地提供不同类别 NGP 的时空动力学信息，有必要提出新的视角和方法。时间序列变化检测方法可以捕捉时空范围内的土地覆盖变化特征（Brown 等人，2020；刘等，2019；Sulla-Menashe 等人，2016；Vogelmann 等人，2016），通常采用单波段定量参数，如归一化差异植被指数（NDVI）、增强植被指数（EVI）等，作为输入数据，而不是多光谱图像。由于历史遥感图像的获取成本大大降低，具有同时相位的时间序列图像（例如，多年的月度或季节性时间序列图像）更方便地用于有效地探索陆地特征的时间动态。目前，时间序列变化检测主要有两种方法。第一种方法涉及使用年度图像来生成跨越多年的时间序列。这种方法的典型算法包括植被变化跟踪器算法（Huang 等人，2010 年）和用于干扰和恢复趋势检测的 LandTrendr 算法（Kennedy 等人，2010 年）。第二种方法首先使用每年所有可用的图像形成年内时间序列，然后将其连接成多年时间序列，并由算法表示，如加性季节和趋势的中断算法（Verbesselt 等人，2010 年）和连续变化检测和分类算法（CCDC）（Zhu&Woodcock，2014 年）。 目前，只有少数研究人员应用时间遥感方法识别耕地 NGP 转化信息（Yu,2020）。
The CCDC algorithm differs from traditional methods as it primarily uses time series from land-based satellites, resulting in less noise. All available Landsat images were employed in this algorithm, providing more comprehensive change detection results compared to using only quasi-annual images, which is more effective in detecting gradual changes (Vogelmann et al., 2016). Currently, CCDC has successfully applied to detect land cover change (Chai & Li, 2023; Xie et al., 2022), forest disturbance (Bullock et al., 2020; Cai et al., 2022), impervious surface expansion (Deng & Zhu, 2020; Xu et al., 2019), and mining disturbance (Xiao et al., 2020). Therefore, this study attempts to use the CCDC algorithm to identify NGP.
CCDC 算法不同于传统方法，因为它主要使用来自陆基卫星的时间序列，从而导致更少的噪声。该算法采用了所有可用的陆地卫星图像，与仅使用准年度图像相比，提供了更全面的变化检测结果，这在检测逐渐变化方面更有效（Vogelmann 等人，2016 年）。目前，CCDC 已成功应用于检测土地覆盖变化（柴&李，2023；谢等，2022），森林扰动（布洛克等，2020；蔡等，2022），不透水表面膨胀（邓&朱，2020；徐等，2019），采矿扰动（肖等，2020）。因此，本研究尝试使用 CCDC 算法来识别 NGP。
To provide long-term NGP evolution characteristics, this study proposes an NGP spatiotemporal change detection method based on Landsat time-series data and the CCDC algorithm, taking Jiashan County in Zhejiang Province as the research area. The specific research objectives were (1) to identify the specific type of NGP based on remote sensing index dynamic trend features; (2) to detect the occurrence time of NGP by CCDC algorithm; and (3) to map the spatiotemporal distribution of NGP.
为了提供长期 NGP 演化特征，本研究以浙江省嘉善县为研究区域，提出了一种基于 Landsat 时间序列数据和 CCDC 算法的 NGP 时空变化检测方法。具体的研究目标是（1）基于遥感指数动态趋势特征识别 NGP 的特定类型；(2)利用 CCDC 算法检测 NGP 的发生时间；（3）绘制 NGP 的时空分布图。
2 MATERIALS 2 材料
2.1 Study area 2.1 研究区域
The study area is in the northeast of Zhejiang province which is an eastern, coastal province of China. Jiashan County is in the center of the Yangtze River Delta, bordering Shanghai to the northeast and Jiangsu province to the north (120°44′ E–121°1′ E, 30°75′ N–31°1′ N) (Figure 1).
研究区域位于中国东部沿海省份浙江省东北部。嘉善县地处长江三角洲中心，东北与上海接壤，北与江苏省接壤（120°44′E-121°1′E，30°75′N-31°1′N）（图 1）。
嘉善县：（1）嘉善在中国的地理位置；（2）2020 年嘉善土地覆盖情况；及(3)2020 嘉善哨兵形象。威利承认，图中的边界受到多重领土主张的影响。[彩色图可在 wileyonlinelibrary.com 查看]
The total area of the county is approximately 507 km2, characterized by dense water networks and flat terrain. The southern part of the county is higher in elevation than the northern part, with an average of about 3.67 m. Jiashan County is an important grain production base of Zhejiang Province, with a larger portion of cultivated land. However, according to relevant data statistics, the area of grain crops planted decreased rapidly from 2010 to 2016, accounting for about 36.56% of the currently cultivated land, indicating the prominent phenomenon of NGP. Meanwhile, due to its advantageous geographical location in the suburban area of Shanghai, it has witnessed rapid development in modern urban agriculture in recent years, resulting in the emergence of characteristic agricultural industries such as yellow peaches, tomatoes, soft-shelled turtles, flowers, and others, leading to a diverse range of NGP types.
全县总面积约 507 km 2 ，水网密布，地势平坦。县城南部海拔高于北部，平均约 3.67 m。嘉善县是浙江省重要的粮食生产基地，耕地比重较大。但据相关数据统计，2010-2016 年粮食作物种植面积快速减少，约占当前耕地的 36.56%，表明 NGP 现象突出。同时，由于其位于上海郊区的优越地理位置，近年来现代都市农业发展迅速，黄桃、番茄、甲鱼、花卉等特色农业产业应运而生，NGP 类型多样。
Such a variable and dynamic geographical environment offer a valuable opportunity to examine our approach to continuous change detection of NGP. The research on the NGP evolution in this region has strong typicality and representativeness, as it can provide valuable insights and guidance for optimizing and adjusting the crop planting structure in economically developed areas.
这种可变和动态的地理环境提供了一个宝贵的机会来检查我们的 NGP 连续变化检测方法。对该地区 NGP 演变的研究具有很强的典型性和代表性，可以为经济发达地区优化调整农作物种植结构提供有价值的见解和指导。
2.2 Definition of NGP types
2.2 NGP 类型的定义
Internationally, there is no clear concept of NGP. Relative researches focus on staple food and cash food production, rice self-sufficiency (Andrade et al., 2022; Haile et al., 2016; Hashmiu et al., 2022), and crop diversity (Gaudin et al., 2015). This concept bears a resemblance to NGP, which involves changing the production patterns on cropland. However, due to varying national circumstances and management objectives, many countries promote crop diversity and the cultivation of economic crops (Nicholson et al., 2021; Renard & Tilman, 2019; Rubhara et al., 2020), while in China, the NGP control has become a government policy.
在国际上，没有明确的 NGP 概念。相关研究侧重于主食和现金食品生产、水稻自给自足（Andrade 等人，2022；海尔等人，2016；Hashmiu 等人，2022 年）和作物多样性（Gaudin 等人，2015 年）。这个概念与 NGP 有相似之处，它涉及改变农田的生产模式。然而，由于不同的国情和管理目标，许多国家促进作物多样性和经济作物的种植（Nicholson 等人，2021；勒纳尔&蒂尔曼，2019；Rubhara 等人，2020），而在中国，NGP 控制已经成为政府政策。
In 2020, the Chinese government issued a document titled “Opinions on preventing non-grain production of cultivated land and stabilizing grain production” (http://www.gov.cn/zhengce/content/2020-11/17/content5562053.htm, accessed on December 22, 2023). The document emphasizes that high-quality cultivated land should be prioritized to produce three major grains: cereals, legumes, and tubers. It strictly prohibits the abandonment and occupation of high-quality cultivated land for forestry, fruit growing, and fishery. For other NGP activities, the document provides no clear regulations, leaving it up to local governments to determine. Due to regional differences, the definition and calculation methods of non-grain NGP in China have been very vague. Broadly, NGP refers to the shift in land use from growing grain to engaging in agricultural activities related to non-grain categories (Cao et al., 2022; Wan, 2023). Existing empirical studies proposing classifications of NGP are primarily based on the actual land use types in the research area (Su et al., 2019). In view of high-resolution remote sensing images in Jiashan and the sown area of different NGP types from Jiashan's statistical Yearbooks, this study mainly focused on 3 categories of NGP, that is, forest, greenhouse, and pond. The forest includes nursery and orchard planting; greenhouse includes all cultivation activities taking place within vegetable greenhouses; and pond includes the transformation of cropland into ponds.
2020 年，中国政府发布了一份名为《关于防止耕地非粮化稳定粮食生产的意见》（http://www.gov.cn/zhengce/content/2020-11/17/content5562053.htm，2023 年 12 月 22 日访问）。文件强调，优质耕地优先生产谷物、豆类、块茎三大粮食。严禁撂荒、占用优质耕地搞林业、种果、渔业。对于 NGP 的其他活动，文件没有做出明确规定，由地方政府自行决定。由于地区差异，我国非粮 NGP 的定义和计算方法一直十分模糊。广义地说，NGP 是指土地利用从种植粮食转向从事与非粮食类别相关的农业活动（曹等人，2022；万，2023）。现有的提出 NGP 分类的实证研究主要基于研究区域的实际土地利用类型（Su 等人，2019）。针对嘉善高分辨率遥感影像和嘉善统计年鉴中不同 NGP 类型的播种面积，本研究主要关注 3 类 NGP，即森林、温室和池塘。森林包括苗圃和果园种植；温室包括在蔬菜温室内进行的所有栽培活动；池塘包括农田变池塘。
2.3 Data acquisition and preprocessing
2.3 数据采集与预处理
2.3.1 Landsat images 2.3.1 陆地卫星图像
This study utilized four Landsat image collections stored on the GEE platform (Gorelick et al., 2017): Landsat4 Thematic Mapper (TM), Landsat5 Thematic Mapper (TM), Landsat7 Enhance Thematic Mapper (ETM+), and Landsat8 Operational Land Imager (OLI) Surface Reflection (SR). We employed all Landsat images within the study area from 1986 to 2022 (USGS http://earthexplorer.usgs.gov/).
这项研究利用了存储在 GEE 平台上的四个 Landsat 图像集（Gorelick 等人，2017 年）：Landsat4 专题制图仪(TM)、Landsat5 专题制图仪(TM)、Landsat7 增强专题制图仪(ETM+)和 Landsat8 业务陆地成像仪(OLI)表面反射(SR)。我们采用了 1986 年至 2022 年研究区域内的所有陆地卫星图像(USGS http://earthexplorer.usgs.gov/)。
The Landsat images used in this study have undergone radiometric correction, terrain correction, and atmospheric correction. Atmospheric correction was performed using the Landsat Ecosystem Disturbance Adaptive Processing System algorithm (Gail et al., 2013). We utilized the QA band of the Landsat SR product to remove pixels that contained clouds, snow, and cloud shadows.
本研究中使用的陆地卫星图像经过了辐射校正、地形校正和大气校正。使用陆地卫星生态系统干扰自适应处理系统算法进行大气校正（Gail 等人，2013 年）。我们利用 Landsat SR 产品的 QA 波段来去除包含云、雪和云阴影的像素。
2.3.2 Ancillary data 2.3.2 辅助数据
In this study, we use several auxiliary data in addition to Landsat. Auxiliary data used in this study included GlobeLand30 (2000, 2010, 2020), GLC_FCS30 (1986–2020, 5a interval), Global Cultivated land (2000–2019, 4a interval), CLCD (1986–2020, annual), CNLUCC (1990, 2000, 2005, 2010, 2015, 2018), and 2022 Sentinel-2 Multispectral Instrument (MSI). Their metadata and references are listed in Table 1 except Sentinel-2 MSI.
在这项研究中，除了陆地卫星之外，我们还使用了几个辅助数据。本研究中使用的辅助数据包括 GlobeLand30（2000、2010、2020）、GLC_FCS30（1986-2020，5a 间隔）、全球耕地（2000-2019，4a 间隔）、CLCD（1986-2020，年度）、CNLUCC（1990、2000、2005、2010、2015、2018）和 2022 Sentinel-2 多光谱仪器（MSI）。除 Sentinel-2 MSI 外，它们的元数据和参考文献列于表 1。
| Product 产品 | Brief description 简要说明 | Temporal characteristics 时间特征 |
Reference 参考 |
|---|---|---|---|
| Global cultivated land 全球耕地 | A machine-learning classification global cropland map with Landsat 基于 Landsat 的机器学习分类全球农田地图 |
2000–2019,4a interval 2000-2019.4 a 区间 | (Potapov et al., 2022) （波塔波夫等人，2022） |
| GlobeLand30 环球 30 | The world's first 30 m land cover product, adopts Pixel and object-based methods with knowledge (POK) 世界上第一个 30 m 土地覆盖产品，采用基于像素和基于对象的知识方法(POK) |
2000, 2010, 2020 | (Chen et al., 2015) （陈等人，2015） |
| The China land cover dataset (CLCD) 中国土地覆盖数据集（CLCD） |
An annual land cover map produced with Google Earth Engine 使用 Google Earth 引擎制作的年度土地覆盖图 |
1990–2022 | (Yang and Huang, 2021) （杨和黄，2021 年） |
| GLC_FCS30 | Constructed through a global spatial–temporal spectra library 通过全球时空光谱库构建 |
1986–2020,5a interval 1986-2020.5 a 区间 | (Zhang et al., 2021) （张等人，2021） |
| CNLUCC | Produced using object-oriented manual visual interpretation with Landsat | 1990, 2000, 2005, 2010, 2015, 2018 | (Xu et al., 2018) |
3 METHODOLOGY 3 方法论
整个方法是在谷歌地球引擎（GEE）平台内实现的。在 GEE 中，我们访问了 Landsat 时间序列数据和 Sentinel 卫星数据，并利用年度合成方法和 CCDC 算法来跟踪 NGP 的时空变化。我们提出了一个 NGP 识别技术框架，由 5 个主要流程组成（图 2）。
- Identification of potential NGP area: We fused multi-source cropland data to create the maximum cropland area and then excluded built-up areas to eliminate the interference of non-agricultural land use changes.
潜在 NGP 面积的识别：我们融合了多源农田数据，以创建最大农田面积，然后排除建成区，以消除非农业土地利用变化的干扰。 - NGP identification based on pixels: First, NDVI, Land Surface Water Index (LSWI), and Plastic Greenhouse Index (PGI) time-series data were constructed by Landsat images from 1986 to 2022. Second, we constructed dynamic trend feature models for three types of NGP based on the mentioned 3 indexes by analyzing the optimal representation of index trajectories under different growth patterns and obtained results at the pixel level. Meanwhile, the CCDC algorithm was used to detect change points in the index trajectories, which allowed us to determine the timing of NGP.
基于像素的 NGP 识别：首先，利用 1986 年至 2022 年的陆地卫星图像构建了 NDVI、陆地地表水指数（LSWI）和塑料温室指数（PGI）时间序列数据。其次，基于上述 3 种指数，构建了不同生长方式下的指数轨迹的最佳表达，并从像素层面上获得了最佳的趋势特征模型。同时利用 CCDC 算法检测指标轨迹中的变点，使我们能够确定 NGP 的运行时间。 - Object-oriented classification based on patches: an object-oriented (OO) Land Use–Land Cover (LULC) classification approach was employed to classify the LULC of Jiashan based on 2022 Sentinel images.
基于斑块的面向对象分类：基于 2022 年哨兵图像，采用面向对象（OO）土地利用-土地覆盖（LULC）分类方法对嘉善的 LULC 进行分类。 - Noise minimization: Combining with the OO classification results, the type of NGP was identified by mode processing inside each patch. A mode filter was used to reduce spatial heterogeneity of temporal results.
噪声最小化：结合 OO 分类结果，通过每个补丁内部的模式处理来识别 NGP 的类型。模式滤波器用于减少时间结果的空间异质性。 -
Map of NGP: Overlaying the classified types and time information, we obtained the spatiotemporal evolution of NGP.
NGP 地图：叠加分类类型和时间信息，我们获得了 NGP 的时空演变。The accuracy of the results was verified using a random stratified sampling method.
采用随机分层抽样方法验证结果的准确性。
提出的识别非粮食生产（NGP）算法的技术框架。LSWI，地表水指数；归一化差异植被指数；PGI，塑料温室指数。[彩色图可在 wileyonlinelibrary.com 查看]
3.1 Identification of potential NGP area
3.1 潜在 NGP 区域的识别
For a comprehensive analysis of the spatial–temporal process of NGP on cultivated land, it is imperative to incorporate all land previously utilized for cultivated land since 1986 in subsequent analysis. CLCD, GlobeLand30, GLC_FCS30, and Global Cultivated land are selected in this study to obtain the potential historical cultivated land extents. Due to imperfections in the land cover and cultivated land products, the potential historical cultivated land extents are determined by extracting the union of cultivated lands from multiple products to minimize the cultivated land omission error. Furthermore, croplands that have been converted to built-up areas are masked using the product GlobeLand30. This allowed us to delineate the potential cropland undergoing NGP in the study area, free from the influence of non-agricultural land use changes. The range of potential NGP land is illustrated in Figure 3(1), where the total potential NGP area is 256.30 km2.
为了全面分析耕地 NGP 的时空过程，必须将 1986 年以来所有以前用于耕地的土地纳入后续分析。本研究选择 CLCD、GlobeLand30、GLC_FCS30 和 Global 耕地来获得潜在的历史耕地范围。由于土地覆盖和耕地产品的不完善性，通过从多个产品中提取耕地的联合来确定潜在的历史耕地范围，以最小化耕地遗漏误差。此外，已经转变为建筑区的农田被产品 GlobeLand30 掩盖。这使我们能够描绘研究区域中正在经历 NGP 的潜在农田，不受非农业土地利用变化的影响。潜在 NGP 土地的范围如图 3（1）所示，其中总潜在 NGP 面积为 256.30 km 2 。
（1）潜在非粮食生产（NGP）面积（浅绿色面积和最大农田抹去 ISA）和（2）NGP 类型识别过程中的像素-斑块转换（A：Goggle earth 中的真实图像；b：OO LULC 分类结果；c：像素级 NGP 类型识别结果；d：模式“2”作为斑块级 NGP 类型）。[彩色图可在 wileyonlinelibrary.com 查看]
3.2 Acquisition of three remote sensing time-series data
3.2 三种遥感时间序列数据的获取
As mentioned in Section 2.2, we categorized the NGP of Jiashan into 3: forest, greenhouse, and pond.
如第 2.2 节所述，我们将嘉善的 NGP 分为 3 类：森林、温室和池塘。
Subsequently, we acquired time-series remote sensing data of the three selected indices, NDVI, LSWI, and PGI, spanning from 1986 to 2022. These data were obtained from multiple Landsat images, including Landsat 4, 5, 7, and 8. The time-series data allowed us to track the temporal dynamics of these indices and analyze the changes in vegetation, water bodies, and plastic greenhouses over the study period in Jiashan County.
随后，我们获取了 1986 年至 2022 年三个选定指数 NDVI、LSWI 和 PGI 的时间序列遥感数据。这些数据是从多个 Landsat 图像中获得的，包括 Landsat 4、5、7 和 8。时间序列数据使我们能够跟踪这些指数的时间动态，并分析嘉善县研究期间植被、水体和塑料温室的变化。
NDVI 是估计特定陆地区域内的绿色植被密度。LSWI 对植被中液态水总量及其土壤背景敏感。PGI（杨等人，2017）放大了塑料温室区域的光谱值，并将其与开阔的农田、裸露的土地和专横的表面积区分开来。较高的 PGI 值表示该区域对应于塑料温室的可能性较高。指数的计算如下：
其中
3.3 Type detection of NGP
3.3 NGP 的类型检测
我们分析了不同增长模式下指数轨迹的最优表示。当农田转为温室时，较低的绿色度和明显的地膜覆盖导致 NDVI 下降和 PGI 增加。退耕还塘后，植被覆盖度明显下降，土壤含水量增加，导致 NDVI 下降，LSWI 增加。在退耕还林过程中，植被覆盖度有明显的增加，反映在 NDVI 指数上呈现明显的增长。同时，由于与粮食作物相比灌溉基础设施较少，土壤含水量下降，导致 LSWI 指数下降。基于这些观察，建立了以下规则来识别不同类型的 NGP（图 4（1））。
- Cultivated Land ≤ Greenhouse: NDVIslope <0&PGIslope >0.
耕地≤温室：NDVI slope <0&PGI slope >0。 - Cultivated Land ≤ Pond: NDVIslope <0&LSWIslope >0.
耕地≤池塘：NDVI slope <0&LSWI slope >0。 - Cultivated Land ≤ Forest: NDVIslope >0&LSWIslope <0.
耕地≤森林：NDVI slope >0&LSWI slope <0。
类型和时间检测：（1）不同 NGP 类型的检测规则和（2）通过连续变化检测和分类算法进行非粮食生产的时间检测。LSWI，地表水指数；归一化差异植被指数；PGI，塑料温室指数。[彩色图可在 wileyonlinelibrary.com 查看]
The median compositing method is used to synthesize available Landsat images taken within a year into a single composite image (Buchner et al., 2020), from which the annual composite values of NDVI, LSWI, and PGI are extracted. Based on the temporal trends and NGP type rules, the specific type of each pixel can be determined, and assigned to each pixel 0 (unchanged), 1 (Forest), 2 (Pond), and 3 (Greenhouse), respectively.
中值合成方法用于将一年内拍摄的可用陆地卫星图像合成为单个合成图像（Buchner 等人，2020 年），从中提取 NDVI、LSWI 和 PGI 的年度合成值。基于时间趋势和 NGP 类型规则，可以确定每个像素的特定类型，并分别为每个像素分配 0（未改变）、1（森林）、2（池塘）和 3（温室）。
3.4 Change detection of NGP
3.4 NGP 的变化检测
The CCDC algorithm utilizes all available Landsat observations for each pixel to construct time series models, which are subsequently used to forecast future observations CCDC (Zhu & Woodcock, 2014). If the values of new consecutive observations diverge from the anticipated range, a break is identified, and a new time series model is created until the next break is detected or all observations are utilized. The detection of a break in the time series models indicates an abrupt change in the land surface environment, which is frequently caused by land use and land cover change. The time series models include Fourier models, which enable the identification of both intra-annual (phenological) and inter-annual (gradual) changes.
CCDC 算法利用每个像素的所有可用陆地卫星观测来构建时间序列模型，随后用于预测未来的观测 CCDC（朱和伍德库克，2014 年）。如果新的连续观测值偏离预期范围，则识别中断，并创建新的时间序列模型，直到检测到下一个中断或利用所有观测值。时间序列模型中断的检测表明地表环境的突然变化，这通常是由土地利用和土地覆盖变化引起的。时间序列模型包括傅立叶模型，其能够识别年内（物候）和年际（渐进）变化。
In this study, we used the CCDC algorithm to identify the last breakpoint, obtaining the timing of NGP (Figure 4(2)).
在这项研究中，我们使用 CCDC 算法来识别最后一个断点，获得 NGP 的时间（图 4（2））。
3.5 Object-oriented classification and noise minimization
3.5 面向对象的分类和噪声最小化
The agricultural pattern in China has been predominantly characterized by smallholder farmers management (Schneibel et al., 2017; Yang, 2021) of which NGP behaviors were organized at the level of individual fields, with consistency in the type of NGP within the same field. From a government management perspective, research on NGP should be conducted at the field level, but the aforementioned NGP type detection model focuses on 30 × 30 m pixel grids, which deviates from the real-world field-level assessment. Meantime, the pixel-based (PB) approach suffers from the “salt-pepper” effect on higher spatial resolutions. An OO LULC classification approach was used to group pixels into meaningful objects based on their spectral, spatial, and temporal characteristics (Petrushevsky et al., 2022).
中国的农业模式主要以小农管理为特征（Schneibel 等人，2017；杨，2021），其中 NGP 行为是在单个领域的水平上组织的，同一领域内 NGP 类型具有一致性。从政府管理角度来看，NGP 的研究应在野外层面进行，但前述 NGP 类型检测模型侧重于 30 × 30 m 像素网格，偏离了真实世界野外层面的评估。同时，基于像素（PB）的方法在较高的空间分辨率上受到“椒盐”效应的影响。OO LULC 分类方法用于根据像素的光谱、空间和时间特征将像素分组为有意义的对象（Petrushevsky 等人，2022）。
This study constructed a method that employs an object-oriented classification approach to process higher-resolution Sentinel data, allowing for the segmentation of regions with consistent land cover types. Subsequently, it compares the results with PB NGP type determinations, and carries out mode processing for NGP types within each segment, ensuring that segments with consistent land cover types correspond to the same NGP type. Figure 3(2) illustrates this process visually.
这项研究构建了一种方法，采用面向对象的分类方法来处理更高分辨率的哨兵数据，允许分割具有一致土地覆盖类型的区域。随后，它将结果与 PB NGP 类型确定进行比较，并对每个区段内的 NGP 类型进行模式处理，确保土地覆盖类型一致的区段对应于相同的 NGP 类型。图 3(2)直观地说明了这一过程。
We used 2022 Sentinel-2 Multispectral Instrument (MSI) for land cover classification in Jiashan County. The Simple Non-Iterative Clustering (SNIC) algorithm is used for super-pixel segmentation, which is available in GEE and has been widely used to group similar pixels and identify spatial clusters (Radhakrishna & Sabine, 2017; Tassi & Vizzari, 2020). We used “Image.Segmentation.seedGrid” function to generate a group of cluster seeds, in which we chose 15 as the superpixel seed location spacing (in pixels). The seed spacing was determined by trial-and-error. SNIC grouped neighboring pixels with similar spectral features, which were based on their NDVI, BSI indexes, and 12 bands median, to create spatially coherent superpixels. Each pixel was assigned to the nearest seed in both spectral and spatial domains, without the need for iterative updates. SNIC requires three parameters in GEE: the “compactness,” which affects the shape of clusters (larger values result in more compact clusters); the “connectivity” parameter, determining whether to consider Rook's or Queen's contiguity when merging adjacent clusters; and a “neighborhood size” to mitigate artifacts at tile boundaries. In this study, 3 parameters were set as follows: compactness = 0, connectivity = 8, and neighborhood size = 256. The output scale of SNIC used the native resolution (S2 = 10).
我们使用 2022 Sentinel-2 多光谱仪器（MSI）对嘉善县进行土地覆盖分类。简单非迭代聚类（SNIC）算法用于超像素分割，该算法在 GEE 中可用，并已广泛用于对相似像素进行分组和识别空间聚类（Radhakrishna&Sabine，2017；Tassi&Vizzari，2020）。我们使用“Image.Segmentation.seedGrid”函数生成一组聚类种子，其中我们选择 15 作为超像素种子位置间距（以像素为单位）。通过试错法确定种子间距。SNIC 根据 NDVI、BSI 指数和 12 波段中值对具有相似光谱特征的相邻像素进行分组，以创建空间相干的超像素。每个像素都被分配给光谱域和空间域中最近的种子，而不需要迭代更新。SNIC 在 GEE 中需要三个参数：“紧凑性”，它影响簇的形状（较大的值导致更紧凑的簇）；“连通性”参数，确定合并相邻簇时是考虑车邻接还是皇后邻接；以及“邻域大小”以减轻图块边界处的伪影。在本研究中，3 个参数设置如下：紧凑性=0，连通性=8，邻域大小=256。SNIC 的输出比例使用原生分辨率（S2=10）。
The mode of pixel NGP types within each patch is utilized to determine the specific type for that patch. This pixel-patch conversion reduced the heterogeneity of NGP types within segments, effectively mitigating the impact of noise. Furthermore, since these patches are morphologically similar to fields (Figure 3(2b)), this approach established a connection between pixels and fields.
利用每个补丁内的像素 NGP 类型的模式来确定该补丁的特定类型。这种像素-斑块转换减少了片段内 NGP 类型的异质性，有效地减轻了噪声的影响。此外，由于这些斑块在形态上类似于场（图 3（2b）），这种方法在像素和场之间建立了联系。
In addition, with a 3 × 3 moving window, a mode filter was used to reduce the spatial heterogeneity of temporal results in Section 3.4.
此外，在第 3.4 节中，对于 3 × 3 移动窗口，使用模式滤波器来减少时间结果的空间异质性。
3.6 Data post-processing and validation
3.6 数据后处理和验证
A random sampling strategy is used to quantify the accuracy of this method. To verify the performance of the multi-index dynamic trend rule in identifying 3 types, 100 samples of each type (300 in total) were selected through Landsat data and Google Earth VHR images. The overall accuracy (OA), kappa coefficient, producer's accuracy (PA), and user's accuracy (UA) were calculated by comparing the results of the sample labels with those from this paper.
随机抽样策略用于量化该方法的准确性。为了验证多指数动态趋势规则在识别 3 种类型中的性能，通过 Landsat 数据和 Google Earth VHR 图像选择了每种类型的 100 个样本（总共 300 个）。通过将样本标签的结果与本文的结果进行比较，计算了总体准确度（OA）、kappa 系数、生产者准确度（PA）和用户准确度（UA）。
To evaluate the accuracy of the CCDC algorithm in identifying the change year, 30 sample points (1100 in total) were selected annually for the time consistency verification based on the method of “stratified random sampling.” We calculated quantitative metrics including OA, Kappa coefficient, PA, UA, and F1score.
为了评估 CCDC 算法在识别变化年份方面的准确性，基于“分层随机抽样”的方法，每年选择 30 个样本点（总共 1100 个）进行时间一致性验证。我们计算了定量指标，包括 OA、Kappa 系数、PA、UA 和 F1 评分。
4 RESULTS 4 结果
4.1 Accuracy assessment 4.1 准确度评估
OA, kappa coefficients, PA, and UA of type detection are shown in Figure 5(1). The OA reached 95.67% and the kappa coefficient was 93.50%, which showed that the multi-index dynamic trend rule successfully classified different types of NGP.
类型检测的 OA、kappa 系数、PA 和 UA 如图 5（1）所示。OA 达到 95.67%，kappa 系数为 93.50%，表明多指标动态趋势规则成功地对不同类型的 NGP 进行了分类。
类型精度和年度精度的验证：（1）不同类型（温室、池塘和森林）的精度，（2）年度精度和年度精度±1 年的总体精度（OA）和 kappa 系数，以及（3）生产者精度（PA）、用户精度（UA）以及年度精度和年度精度±1 年的 F1 比较。[彩色图可在 wileyonlinelibrary.com 查看]
Figure 5(2,3) showed the OA, kappa coefficients, PA, UA, and F1score of time detection. The OA was 85.26% and the kappa coefficient was 84.83%. Despite the overall high OA value, significant variations in PA exist across different years. The accuracy ranged from a minimum of 50% in 2012 to a maximum of 100% in 2018. This variation could be attributed to the fact that the conversion of NGP took more than 1 year and happened during the intervals between years. If the yearly accuracy was relaxed to ±1 year, UA, PA, and F1 improved significantly (Figure 5(3)), with the lowest PA would reach 86.67% and UA would reach 83.30%.
图 5（2，3）显示了时间检测的 OA、kappa 系数、PA、UA 和 F1 评分。OA 为 85.26%，kappa 系数为 84.83%。尽管总体 OA 值较高，但不同年份的 PA 存在显著差异。准确率从 2012 年的最低 50%到 2018 年的最高 100%不等。这种变化可能归因于 NGP 的转换需要 1 年以上的时间，并且发生在年份之间的间隔。如果将年度准确度放宽至±1 年，UA、PA 和 F1 显著提高（图 5（3）），PA 最低将达到 86.67%，UA 将达到 83.30%。
4.2 The statistical characteristics of NGP
4.2 NGP 的统计特征
Overall, the total amount of NGP in Jiashan exhibited a continued increasing trend from 1986 to 2022, as shown in Table 2. During the past 37 years, the cumulative percentage of NGP (total NGP area to potential NGP area) increased from 0.02% to 20.69%, with a net increase of 5304.06 ha. The total area of cultivated land changed to greenhouse, pond, and forest was 3113.03, 1612.83, and 578.19 ha, accounting for 58.69%, 30.41%, and 10.90% of their total area in 2022, respectively.
总体而言，嘉善 NGP 总量从 1986 年到 2022 年呈持续增长趋势，如表 2 所示。在过去的 37 年里，NGP 的累计百分比（NGP 总面积与潜在 NGP 面积）从 0.02%增加到 20.69%，净增加 5304.06 公顷。2022 年，转为温室、池塘和森林的耕地总面积分别为 3113.03、1612.83 和 578.19 公顷，分别占其总面积的 58.69%、30.41%和 10.90%。
| Year 年 | Greenhouse (ha) 温室（公顷） | Pond (ha) 池塘（公顷） | Forest (ha) 森林（公顷） | Total (ha) 总计（公顷） | Cumulative percentage (%) 累计百分比（%） |
|---|---|---|---|---|---|
| 1986 | 2.81 | 1.03 | 0.35 | 4.19 | 0.02 |
| 1995 | 414.44 | 124.64 | 80.94 | 620.01 | 2.42 |
| 2005 | 885.58 | 753.72 | 177.27 | 1816.57 | 7.09 |
| 2015 | 1957.95 | 1194.57 | 421.76 | 3574.28 | 13.95 |
| 2022 | 3113.03 | 1612.83 | 578.19 | 5304.06 | 20.69 |
Interannual variability and trends of NGP are shown in Figure 6(1). Before 2000, the NGP annual variation increased slowly at an average annual change of approximately 79.37 ha. For the rest of the period, it demonstrated a continuous v-shaped change, with the highest point being 384.93 ha in 2013. After 2012, the NGP annual variation area was significantly larger than the previous time period, which shows that the NGP in Jiashan was becoming more severe during these years.
NGP 的年际变化和趋势如图 6（1）所示。2000 年以前，NGP 年变化缓慢增加，平均年变化约为 79.37 ha。在此期间的其余时间里，它呈现出连续的 v 形变化，最高点是 2013 年的 384.93 公顷。2012 年以后，NGP 年变化面积明显大于前一个时间段，这表明嘉善的 NGP 在这些年变得更加严重。
非粮食生产（NGP）年际变化和趋势：（1）1986 年至 2022 年年变化面积和 NGP 累计百分比的变化，以及（2）按类别年变化面积饼图。[彩色图可在 wileyonlinelibrary.com 查看]
The annual variation area changes and the spatial distribution of different NGP categories over time exhibited significant inconsistencies (Figures 6(2) and 7).
不同 NGP 类别随时间的年变化面积变化和空间分布表现出显著的不一致性（图 6（2）和 7）。
1986 年至 2022 年非粮食生产的时空变化：（1）温室，（2）池塘，（3）森林。[彩色图可在 wileyonlinelibrary.com 查看]
Greenhouses are the most prominent type reaching its peak at 243.11 ha in 2017, which was concentrated in the northeastern part of the main county town of Jiashan and the southern Ma Jiaqiao village, with some areas distributed in other villages by 2022. According to the information, Majiaqiao Village currently has around 3000 acres of facility greenhouses, making it renowned as the “Hometown of Chinese Sweet Melons.” Jiashan County is in a typical urban suburban area, about 50 km away from Shanghai. Freshly picked fruits and vegetables are sold to Shanghai every day, earning Jiashan the reputation of being Shanghai's “off-campus vegetable basket.” With the high demand for fruits and vegetables and good agricultural resources in Jiashan, Greenhouses have become a prominent feature of Jiashan's distinctive agriculture and a crucial avenue for increasing farmers' income.
温室是最突出的类型，2017 年达到峰值 243.11 公顷，集中在嘉善主要县城的东北部和南部的马家桥村，到 2022 年部分地区分布在其他村庄。资料显示，马家桥村现有设施大棚 3000 亩左右，被誉为“中国甜瓜之乡”。嘉善县位于典型的城市郊区，距上海约 50 公里。每天新鲜采摘的果蔬销往上海，为嘉善赢得了上海“校外菜篮子”的美誉。在嘉善果蔬需求量大、农业资源好的情况下，温室大棚成为嘉善特色农业的突出特色和农民增收至关重要的大道。
Pond emerged as the predominant type between 1997 and 2002 due to the extensive utilization of shoals and low-lying fields for fish pond construction in Jiashan in the early 20th, predominantly concentrated around rivers and lakes. Pond farming has long been a key industry in Jiashan County. For farmers, the pond is also an important way to adjust the agricultural structure. In addition, this type is gradually concentrated in the surrounding areas of the northern Fen Lake, closely tied to the development of the leisure and sightseeing agricultural belt. Utilizing the picturesque natural landscapes of Fen Lake, the development of activities such as fishing, sightseeing, and dining will inevitably occupy a substantial amount of cropland.
由于 20 世纪 20 年代初嘉善广泛利用浅滩和低洼地建设鱼塘，池塘在 1997 年至 2002 年间成为主要类型，主要集中在河流和湖泊周围。池塘养殖长期以来一直是嘉善县的重点产业。对农民来说，池塘也是农业结构调整的重要途径。此外，该类型逐渐集中在北部汾湖周边地区，与休闲观光农业带的发展紧密相连。利用汾湖风景如画的自然景观，发展垂钓、观光、餐饮等活动，必然会占用大量耕地。
The forest is the smallest proportion of 3 types. The annual change of which exhibited a marked increase after 2008 with the average area being 25.31 ha compared to 9.02 ha before and the maximum value of 51.97 ha was recorded in 2013. These areas were primarily clustered along the Shanghai-Kunming Expressway (opened to traffic on October 26, 2010). With the increase in urban greenery demand and the upgrading of the diet structure of urban and rural residents, farmers have gradually chosen this conversion to achieve higher economic profits. The distribution pattern near the Shanghai-Kunming Expressway underscores the significance of transportation in forestry and fruit growing, which not only reduces transportation costs but also facilitates pick-your-own tourism.
森林是 3 种类型中比例最小的。2008 年后，其年度变化显著增加，平均面积为 25.31 公顷，而之前为 9.02 公顷，2013 年记录的最大值为 51.97 公顷。该等区域主要集中于沪昆高速公路（于二零一零年十月二十六日通车）沿线。随着城市绿化需求的增加和城乡居民饮食结构的升级，农民逐渐选择了这种转换以获得更高的经济利润。靠近沪昆高速公路的分布格局凸显了运输在林果种植中的重要性，既降低了运输成本，又方便了自采旅游。
4.3 Spatial autocorrelation characteristics of NGP
4.3 NGP 的空间自相关特性
To demonstrate the spatial expansion of NGP, the total research period was divided into 4 periods, namely 1986–1995, 1995–2005, 2005–2015, and 2015–2022. The spatial autocorrelation characteristics of different NGP categories during these periods were analyzed in terms of global Moran's I (Figure 8) and the local Moran's I (Figure 9).
为了证明 NGP 的空间扩展，将整个研究期分为 4 个时期，即 1986-1995 年、1995-2005 年、2005-2015 年和 2015-2022 年。根据全局 Moran I（图 8）和局部 Moran I（图 9）分析了这些时期不同 NGP 类别的空间自相关特征。
按类别划分的不同时期的全球莫兰 I。[彩色图可在 wileyonlinelibrary.com 查看]
空间关联图的局部指标。[彩色图可在 wileyonlinelibrary.com 查看]
The Global Moran's I index is used to assess whether there is statistically significant clustering or dispersion of NGP overall. The value of Global Moran's I index ranges from −1 to 1, where positive values indicate clustering, negative values indicate dispersion, and a value of zero indicates random distribution. Local spatial autocorrelation is used to explore the spatial clustering characteristics of high and low values of NGP at local spatial locations.
全球莫兰 I 指数用于评估 NGP 总体上是否存在统计学上显著的聚类或分散。全局莫兰 I 指数的值范围为-1 至 1，其中正值表示聚类，负值表示分散，零值表示随机分布。局部空间自相关用于探索 NGP 在局部空间位置的高值和低值的空间聚类特征。
In general, except for the “pond” category from 2005 to 2015 and the “forest” category from 2015 to 2022, all other categories of NGP exhibited spatial dependence, as indicated by global Moran's I value ranging from 0.1148 to 0.3429 during these periods. The Global Moran's I index for greenhouse displayed a V-shaped pattern of variation, while the pond and plant forest exhibited an inverted V-shaped pattern.
总的来说，除了 2005 年至 2015 年的“池塘”类别和 2015 年至 2022 年的“森林”类别外，所有其他类别的 NGP 都表现出空间依赖性，如这些期间全球莫兰 I 值在 0.1148 至 0.3429 之间所示。全球温室 Moran I 指数呈 V 型变化，而池塘和植物林呈倒 V 型变化。
To better demonstrate the specific spatial development process of NGP, the local indicators of spatial association maps of each type were also prepared. Different types indicated conspicuous regional disparities. Most HH regions of a greenhouse are situated primarily in the eastern region, with some additional concentrations in the southwestern region. Pond sites exhibited a concentration in the northern region, particularly around lakes and rivers. Plant nursery and orchard hotpots were observed in the southeastern of Jiashan. The three categories of HH regions exhibit mutual exclusivity, with no overlapping areas observed among them. This spatial pattern suggests that the emergence of a dominant NGP type may exert an inhibitory effect on other categories, thereby shaping the unique spatial distribution observed in the study area.
为了更好地展示 NGP 的具体空间发展过程，还编制了各类型空间关联图的局部指标。不同的类型表明明显的地区差异。温室的大多数 HH 区域主要位于东部地区，在西南部地区也有一些额外的集中。池塘遗址集中在北部地区，特别是湖泊和河流周围。嘉善东南部观察到苗圃和果园火锅。三类 HH 区域表现出互斥性，在它们之间没有观察到重叠区域。这种空间模式表明，一种优势 NGP 类型的出现可能会对其他类别产生抑制作用，从而塑造在研究区域观察到的独特空间分布。
5 DISCUSSION 5 讨论
5.1 Advantages of this method of multi-index dynamic trend rule using time-series data
5.1 利用时间序列数据的多指标动态趋势规则的这种方法的优点
Monitoring the dynamic shifts in human-dominated or -induced cultivated land use is essential for ensuring food security and using remote sensing to track long-term cultivated land changes is a primary research focus because of its difficulty (Xu et al., 2018). The sown area in China's statistical yearbook focuses on NGP on all land, breaking away from the scope of cultivated land and failing to consider the actual changes in cultivated land. Neither the spatial distribution of cultivated land extent nor the temporal consistency of inter-annual statistics is ensured (Xu et al., 2017). The Landcover classification method could not capture the exact times of cultivated land use shifts between the time periods. Image noise and classification algorithms can cause uncertainty in the classification results of each period which is amplified in the trajectory of land cover change. Therefore, it is challenging to ascertain whether the observed changes are actual or caused by image misclassification during a particular time period. The current cultivated land data products, such as CLCD, only offer information on the dynamics of cultivated land, and do not capture any changes in production methods within cultivated land.
监测人类主导或诱导的耕地利用的动态变化对于确保粮食安全至关重要，使用遥感跟踪长期耕地变化是一个主要的研究重点，因为它很困难（Xu 等人，2018）。我国统计年鉴中的播种面积关注的是所有土地上的 NGP，脱离了耕地范围，没有考虑耕地的实际变化。既不能保证耕地范围的空间分布，也不能保证年际统计的时间一致性（徐等，2017）。土地覆盖分类方法无法捕捉不同时间段之间耕地利用变化的确切时间。图像噪声和分类算法会导致各时期分类结果的不确定性，这种不确定性在土地覆盖变化轨迹中被放大。因此，确定观察到的变化是实际的还是由特定时间段期间的图像错误分类引起的是具有挑战性的。目前的耕地数据产品，如 CLCD，仅提供耕地动态信息，并不能捕捉耕地内部生产方式的任何变化。
In this study, we construct rules for different types of NGP from the perspective of typical remote sensing index changes caused by NGP, then incorporate these rules in extracting dynamic change and we adopt the CCDC algorithm for the time of different NGP types detection using dense data stacks from Landsat 4, 5, 7, and 8. This algorithm utilizes all available Landsat observations for each pixel to construct time series models. It is worth noting that using the CCDC algorithm alone cannot classify the types of NGP, which is because the outputs of CCDC algorithms are a binary result of change and non-change classifications with magnitude and direction. Establishing NGP decision rules is necessary and critical.
在本研究中，我们从 NGP 引起的典型遥感指数变化的角度为不同类型的 NGP 构建规则，然后将这些规则纳入动态变化提取中，并采用 CCDC 算法使用来自 Landsat 4、5、7 和 8 的密集数据堆栈对不同 NGP 类型进行检测。该算法利用每个像素的所有可用陆地卫星观测来构建时间序列模型。值得注意的是，单独使用 CCDC 算法无法对 NGP 的类型进行分类，这是因为 CCDC 算法的输出是具有大小和方向的变化和非变化分类的二元结果。建立 NGP 决策规则是必要和关键的。
To achieve the connection between pixels and fields for better government management and minimize the generation of “salt and pepper” noise, we used an OO LULC classification method for the 2022 Sentinel-2 composite image, to obtain a group of patches which is composed of several geographically adjacent pixels. Determine the NGP type of each patch by taking the mode of pixel type within it and assigning these patches again with new categories.
5.2 Compared to the tradition method using CNLUCC
CNLUCC is a national-scale, multi-temporal LUCC database of China which is constructed through object-oriented manual visual interpretation with Landsat as its primary information source. We obtained the potential NGP region by extracting the cultivated land from CNLUCC dataset in 1990. Limited by the fact that CNLUCC does not capture the change of greenhouse and the conversion of cultivated land into the forest is relatively insignificant, the comparison solely focused on the “pond” type. It should be stated that CNLUCC is not a true reference and cannot be regarded as a factual basis. Nevertheless, the comparison enabled us to recognize the synergy level between our method and the traditional landcover classification comparison method.
Figure 10 shows the spatial and temporal distribution of NGP with two different methods and compares their spatiotemporal consistency. The temporal results obtained from CNLUCC are consistent with the trends observed in this study, with the maximum change occurring between 1995 and 2000 (Figure 10(1)). However, our study showed better consistency with the actual situation compared to CNLUCC when compared with Google Earth images from 2018. Specifically, CNLUCC exhibited noticeable geographical displacements (Figure 10(2a)), detected many pseudo changes instead of true changes (Figure 10(2b)) and failed to recognize areas where changes had occurred (Figure 10(2c)).
5.3 Policy implications
Based on the analysis of remote sensing data, in-depth research can be conducted to identify high-risk areas and key factors and formulate tailored governance measures and policies for different regions. The actual situation of long-term NGP on cultivated lands should be considered when carrying out prohibition policies. For example, in this case study, we observed that the 1816.57 ha area had a history of NGP for more than about 20 years, which may indicate the industrialization of NGP. Directly uprooting crops on cropland without considering the reality of long-term cultivation of non-grain crops, leading to a return to poverty among farmers, is not in line with the laws of development. In addition, it is not advisable to include the development of fruit growing and forestry on sloping cultivated land with poor basic tillage conditions in NGP. Planting economic crops on these lands superficially violates land policies, but in fact conforms to development regularities. These efforts are directed toward formulating local policies that are better tailored and responsive.
Utilize remote sensing data technology for regular monitoring and evaluation of NGP and establish a robust remote sensing monitoring system to obtain real-time data. This will provide policymakers with detailed and up-to-date data on the dynamics of NGP, enabling them to formulate effective policies for farmland management and protection. It is worth noting that a single satellite image is insufficient to judge the illegality of cropland use changes, as seasonal variations in vegetation can lead to false positives, such as waterlogged paddy fields appearing to be water bodies. Time series monitoring can effectively address this issue by providing more data points to distinguish between similar objects. Therefore, for detecting changes in land use, it is important to use time-series remote sensing as a starting point, followed by field surveys to verify the findings and make objective assessments of the legality of land use changes based on evidence and policy.
Optimize food security strategy and land use planning to strike a delicate equilibrium between non-grain industries and food security and ecological conservation. When formulating cropland protection goals, NGP behaviors that do not damage the soil layer, allowing for rapid restoration of the cultivation of staple crops, such as the cultivation of oilseeds and vegetables, should not be categorically prohibited. When devising land use plans, it is essential to duly account for the developmental needs of non-grain industries and devise strategies for the judicious utilization of land resources. In addition, the current boundaries of permanent basic farmland in current land planning mostly rely on the previous-stage plans, leading to situations that are not in line with reality. Due to the lack of awareness regarding policies related to permanent basic farmland, many farmers engage in NGP without being adequately informed. Integrate with land use planning to clearly define the boundaries of permanent basic farmland and general croplands, thereby avoiding “passive” non-grain conversion activities.
5.4 Limitations
Although the performance of NGP mapping was very acceptable, with over 85% overall accuracy, it should be noted that the method still has limitations. Specifically, it cannot discern repetitive changes, such as the seasonal rotation between food and non-food crops. Furthermore, it may not effectively distinguish between the conversion of cultivated land to forest and cropland abandonment. These challenges underscore the need for complementary approaches in characterizing and monitoring the full spectrum of NGP dynamics.
6 CONCLUSION
Food security is seriously threatened by NGP in China. This study proposes a method for detecting the spatiotemporal evolution of NGP with Landsat time series-data. The application in Jiashan proved the applicability of the method. The accuracy demonstrates that this method can effectively detect planting structure adjustment on cropland and provide information on the type and time of changes.
NGP can be attributed to the progress in social and economic development. Farmers have made a deliberate choice to cultivate cash crops over traditional grain crops, following the market's law of seeking optimal profits. Our results suggest that it is necessary to revisit traditional NGP identification methods and that Landsat time series change detection can make an important contribution to the monitoring of NGP change. Moreover, the method in this paper was performed on the GEE platform, which can be flexibly transplanted to different spatial and temporal scales, providing a new research idea for the detection of human-dominated or -induced long-term cultivated land changes involving NGP.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflicts of interest.
