城市边界线预测（根据灯光指数）(PUL)

1.EXEALL.m

function EXEALL(FilePath, FileName)
%执行所有流程
% FilePath: 文件夹所在路径
% FileName: 文件夹名称

FullPath = [FilePath , FileName , '\' , FileName , '.txt'];
allPixelPath = [FilePath , FileName , '\', FileName , '.allpixel.csv'];
TestdataPath = [FilePath , FileName , '\data\Urban.Test.csv'];
TraindataPath = [FilePath , FileName , '\data\Urban.Train.csv'];
netPath = [FilePath , FileName , '\net\Urban.net.mat'];
predictFld = [FilePath , FileName , '\predict\'];
accFld = [FilePath , FileName , '\acc\'];
FileFld = [FilePath , FileName , '\'];

if ~exist(allPixelPath,'file')
allpixel(FullPath,allPixelPath);
end
if ~exist(TraindataPath,'file')
FormatData(allPixelPath, TestdataPath, TraindataPath,4000,5000,1000);
end
if exist(TraindataPath,'file')
pbl_train(TraindataPath,netPath);
end
if exist(netPath,'file')
TestDataPredict(TestdataPath,netPath,predictFld,FileName);
TestDataAcc(TestdataPath,predictFld,accFld,FileName);
PredictAll(allPixelPath,netPath,FileFld,FileName)
end
end

2.

Mapping Regional Urban Extent Using NPP-VIIRS DNB and MODIS NDVI Data

利用NPP-VIIRS dnb和modis ndvi数据绘制区域城市范围图

Abstract: The accurate and timely monitoring of regional urban extent is helpful for analyzing urban sprawl and studying environmental issues related to urbanization. This paper proposes a classification scheme for large-scale urban extent mapping by combining the Day/Night Band of the Visible Infrared Imaging Radiometer Suite on the Suomi National Polar-orbiting Partnership Satellite (NPP-VIIRS DNB) and the Normalized Difference Vegetation Index from the Moderate Resolution Imaging Spectroradiometer products (MODIS NDVI). A Back Propagation (BP) neural network based one-class classification method, the Present-Unlabeled Learning (PUL) algorithm,is employed to classify images into urban and non-urban areas. Experiments are conducted in mainland China (excluding surrounding islands) to detect urban areas in 2012. Results show that the proposed model can successfully map urban area with a kappa of 0.842 on the pixel level. Most of the urban areas are identified with a producer’s accuracy of 79.63%, and only 10.42% the generated urban areas are misclassified with a user’s accuracy of 89.58%. At the city level, among 647 cities, only four county-level cities are omitted. To evaluate the effectiveness of the proposed scheme, three contrastive analyses are conducted: (1) comparing the urban map obtained in this paper with that generated by
the Defense Meteorological Satellite Program/Operational Linescan System Nighttime Light Data (DMSP/OLS NLD) and MODIS NDVI and with that extracted from MCD12Q1 in MODIS products;(2) comparing the performance of the integration of NPP-VIIRS DNB and MODIS NDVI with single input data; and (3) comparing the classification method used in this paper (PUL) with a linear method(Large-scale Impervious Surface Index (LISI)). According to our analyses, the proposed classification scheme shows great potential to map regional urban extents in an effective and efficient manner.
Keywords: urban mapping; one-class; NPP-VIIRS DNB; MODIS NDVI; large scale

摘要：准确、及时地监测区域城市规模，有助于分析城市蔓延，研究与城市化有关的环境问题。本文提出了一种大规模城市范围图的分类方案，通过结合Suomi国家极地轨道伙伴卫星可见红外成像辐射计套件的日/夜波段（NPP-VIIRS DNB）和中分辨率成像光谱辐射计产品中的归一化差异植被指数（MODIS NDVI）。采用基于BP神经网络的一类分类方法（当前无标记学习算法，PUL），将图像分为城市和非城市两类。2012年，中国大陆（不包括周边岛屿）进行了城市区域检测实验。结果表明，该模型能在像素级上绘制出卡帕为0.842的城区地图。大多数城市地区被确定为生产者的准确度为79.63%，只有10.42%的城市地区被错误分类，用户的准确度为89.58%。在城市层面，647个城市中，只有4个县级城市被省略。为了评估拟议方案的有效性，进行了三次对比分析：(1)将本文获得的城市地图与国防气象卫星计划/业务线扫描系统夜间光照数据（DMSP/OLS NLD）和MODIS NDVI生成的城市地图以及从MCD12Q1中提取的MODIS产品进行比较；（2）比较NPP-VIIRS DNB和MODIS NDVI与单输入数据的集成性能；(3) 将本文所用的分类方法（pul）与线性方法（大型不透水面指数（lisi））进行比较。根据我们的分析，提出的分类方案显示出以有效和高效的方式绘制区域城市范围图的巨大潜力。
关键词：城市地图；一类；NPP-VIIRS DNB；MODIS NDVI；大规模

1. Introduction
Urban land occupies a relatively small fraction of the Earth’s surface, but is the main area of human activities [1]. Rapid socio-economic development and population growth have greatly encouraged the expansion of urban land areas [2,3]. According to the World Bank, by the end of 2015, the global urbanization rate had hit 53.85%, a dramatic increase from 46.54% in 2000. If current trends continue, by 2030, urban land cover will grow by 1.2 million km2, nearly tripling the global urban land area of circa 2000 [4]. This increase has a high-probability of occurring in developing countries with a rapid pace of urbanization, such as China. Although urbanization is an effective promotion of social progress, urban expansion has inevitably caused numerous environmental problems, resulting in natural habitat loss, threatening biodiversity, affecting local climate, etc. [5–11]. It is essential to obtain timely and accurate information on urbanized areas to monitor the environmental and socioeconomic processes [12,13].

城市土地只占地球表面的一小部分，但却是人类活动的主要区域。快速的社会经济发展和人口增长极大地促进了城市用地面积的扩大。根据世界银行的数据，到2015年底，全球城市化率已从2000年的46.54%大幅上升至53.85%。如果目前的趋势继续下去，到2030年，城市土地覆盖面积将增加120万平方公里，几乎是2000年全球城市土地面积的三倍。这种增长很可能发生在中国等城市化速度较快的发展中国家。虽然城市化是社会进步的有效促进，但城市扩张不可避免地造成了许多环境问题，导致自然栖息地丧失、生物多样性受到威胁、影响当地气候等。对环境和社会经济过程进行监测，必须及时准确地获得有关城市化地区的信息。

Land cover maps derived from remotely-sensed data are a valuable source for mapping urban area and monitoring urban dynamics [14–19]. Images with high or medium spatial resolution are commonly used for characterizing the urban structures due to their detailed spatial observations of cities [20–22]. However, the coverage range of one scene in a high or medium spatial resolution image is limited. For example, IKONOS collects high-resolution imagery at 1 and 4 m resolution, and the swath of a single scene is 11 km × 11 km. When applied at the regional scale (e.g., China, covering approximately 9.6 million km2), a large number of images is needed to cover the entire area. It is time consuming and labor intensive to process multi-fold pixels [23]. High- and medium-spatial resolution images are also frequently affected by cloud cover. It is challenging to select high-quality images around the same time for a large area. Additionally, the sensors collecting high- and medium-resolution images have a relatively long revisiting period. For example, Landsat sensors, providing medium-resolution imagery of 30 m, have a revisit rate at 16 days, which reduces the number of suitable images. Thus, at the national or continental scale, coarse imagery with wider swaths and higher revisit rates is the preferred imagery for macroscopically monitoring urban extents [24]

从遥感数据中获得的土地覆盖图是绘制城区地图和监测城市动态的重要来源[14–19]。由于城市的详细空间观测，高或中等空间分辨率的图像通常用于描述城市结构[20-22]。然而，在高分辨率或中分辨率图像中，一个场景的覆盖范围是有限的。例如，ikonos以1米和4米的分辨率收集高分辨率图像，单个场景的宽度为11公里×11公里。当以区域尺度（如中国，面积约960万平方公里）应用时，需要大量图像覆盖整个区域。处理多倍像素非常耗时和费力[23]。高分辨率和中分辨率图像也经常受到云量的影响。在大范围内同时选择高质量的图像是一个挑战。此外，收集高分辨率和中分辨率图像的传感器具有相对较长的回访周期。例如，提供30米中等分辨率图像的陆地卫星传感器在16天内有一个重访率，这减少了合适图像的数量。因此，在国家或大陆尺度上，具有更宽范围和更高重新访问率的粗糙图像是宏观监测城市范围的首选图像[24]

As the main data source of coarse images, the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument can provide global imagery (500 to 1000 m spatial resolution) with 1–2-day temporal resolution. The dimensions of one scene are 2400 × 2400 rows/columns (about 1.4 million km2). Many researchers have explored the capability of MODIS reflectance data in mapping urban areas [12,18,25]. The Normalized Difference Vegetation Index (NDVI) products derived from reflectance data are frequently used to rapidly characterize urban extent [26–29]. Since urban areas are dominated by impervious surfaces, the vegetation distributions inside and outside cities have obvious differences, which can be reflected in the NDVI [30]. Multi-temporal or time series NDVI are often adopted to remove the cloud contamination, which always occurs in a single time image [31]. However, either MODIS reflectance data or the derived NDVI alone can reflect the physical properties of the land surface. Urban physical properties are similar to those of non- or low-vegetation land covers, and some non-urban blocks may have similar reflecting spectral curves as urban blocks, such as bare soils. Besides, urban areas can be regarded as the union of multi-features (e.g., trees, asphalt road, buildings, etc.), but the specific composition of each city is different, resulting in urban spectral features varying from city to city. There is no universal standard to distinguish city from the background. Particularly, pixels in coarse images cover a larger ground surface and contain more features. The spectral heterogeneity and homogeneity of urban land make it easily confused with other land covers, resulting in biased estimated (overestimated or underestimated). Thus, mapping urban extent from physical information alone is challenging.

中分辨率成像分光辐射计（modis）作为粗图像的主要数据源，可以提供具有1-2天时间分辨率的全球图像（500-1000米空间分辨率）。一个场景的尺寸为2400×2400行/列（约140万平方公里）。许多研究人员已经探索了modis反射数据在绘制城市区域图方面的能力[12,18,25]。从反射率数据中得到的归一化差异植被指数（ndvi）产品经常被用来快速表征城市范围[26-29]。由于城市区域以不透水面为主，城市内外植被分布存在明显差异，可以在NDVI中反映出来[30]。多时间或时间序列的NDVI通常被用来消除云污染，而云污染总是发生在一张单一的时间图像中[31]。然而，无论是modis反射率数据还是仅导出的ndvi都可以反映地表的物理性质。城市物理性质与非植被或低植被土地覆盖相似，一些非城市块体可能具有与城市块体（如裸地）相似的反射光谱曲线。此外，城市区域可以看作是多特征（如树木、柏油路、建筑等）的结合体，但每个城市的具体组成是不同的，导致城市的光谱特征因城市而异。没有通用的标准来区分城市和背景。特别是，粗糙图像中的像素覆盖较大的地面，并包含更多的特征。城市土地的光谱异质性和同质性使其很容易与其他土地覆盖相混淆，从而导致有偏估计（高估或低估）。因此，仅从物理信息中绘制城市范围是一项挑战。

Another coarse data source, the nighttime lighting data, regarded as a sign of human activities,can provide different urban information from spectral data. It can distinguish human urban areas with artificial lights from the dark background at night [32] and has been employed to estimate economic conditions and energy consumption [33,34]. The most widely-used nighttime light data are the stable Nighttime Light Data on the Defense Meteorological Satellite Program/Operational Linescan System (DMSP/OLS NLD) [35–39], but they have some limitations related to their coarse spatial resolution (30 arc seconds, 5 km × 5 km at nadir), blooming effects (the dispersion of light from built-up areas into non-light areas), saturation in urban areas and intra-sensor calibration problems, resulting in misestimates of urban land [40,41]. Some studies have found that the combined use of spectral data and nighttime light data can reduce the blooming effect and pixel saturation of nighttime light products, as well as provide better performance than using individual nighttime light data [42,43]. However, this integration, which provides only supplemental information for DMSP/OLS NLD, cannot eliminate data limitations, i.e., coarse spatial resolution of about 1 km and small data ranges from 0 to 63.

另一个粗略的数据来源，夜间照明数据，被认为是人类活动的标志，可以提供不同于光谱数据的城市信息。它可以将人工照明的城市区域与夜间的黑暗背景区分开[32]，并被用于估算经济条件和能源消耗[33,34]。最广泛使用的夜间光数据是国防气象卫星计划/作战线扫描系统（DMSP/OLS NLD）上稳定的夜间光数据[35–39]，但它们存在一些局限性，这些局限性与它们的粗糙空间分辨率（30弧秒，5 km×5 km在最低点）、开花效应（来自内置U-U的光的分散）有关。p区域变为非光区），城市区域饱和和传感器内部校准问题，导致对城市土地的错误估计[40,41]。一些研究发现，将光谱数据和夜间光数据结合使用，可以降低夜间光产品的晕染效应和像素饱和度，并提供比单独夜间光数据更好的性能[42,43]。但是，这种仅为DMSP/OLS NLD提供补充信息的集成无法消除数据限制，即大约1公里的粗空间分辨率和0到63之间的小数据范围。

The next generation of nighttime light observations, the Day/Night Band included in the Visible Infrared Imaging Radiometer Suite on the Suomi National Polar-orbiting Partnership Satellite (NPP-VIIRS DNB) [44], has made dramatic improvements to the spatial, spectral and radiometric resolution. The increased spatial resolution (15 arc seconds, 742 m × 742 m), lower light imaging detection limits (~2 × 10−11 W·cm−2·sr−1) and higher radiometric quantization (14 bit) allow NPP-VIIRS DNB to provide more details of the city, bringing new insight into extracting urban extent [41,45]. However, nighttime light data are seriously influenced by different economic conditions [46,47]. Using solely NPP-VIIRS DNB may produce inaccurate spatial patterns of cities, especially for considerably different economic conditions. Concerning this issue, Guo et al. [31] established a Large-scale Impervious Surface Index (LISI) integrating NPP-VIIRS DNB and MODIS NDVI variables, which greatly improved the extraction performance. The LISI-based approach provides accurate spatial patterns from high values in core urban areas to low values in rural areas, with an overall root mean squared error of 0.11. Similarly, Sharma et al. [48] proposed an Urban Built-up Index (UBI) combining the MODIS multispectral imagery with NPP-VIIRS DNB to generate a 500-m resolution map of urban built-up areas at the global scale. The combination of spectral data and VIIRS nighttime light data significantly reduces the misclassification errors, capturing more detailed information of urban spatial patterns. Nevertheless, indexes only estimate the performance in the linear model. More research is still required to explore the nonparametric algorithms and broaden the joint utilization of multi-source coarser images for urban land extraction [49].

下一代夜间光观测，包括在Suomi国家极地轨道伙伴关系卫星（NPP-VIIRS DNB）可见红外成像辐射计套件中的白天/夜间波段，已经对空间、光谱和辐射分辨率作出了巨大的改进。增加的空间分辨率（15弧秒，742米×742米），较低的光成像检测限（~2×10−11 W·cm−2·sr−1）和较高的辐射量化（14位），使得NPP-VIIRS DNB能够提供更多的城市细节，为提取城市范围带来新的见解[41,45]。然而，夜间光照数据受到不同经济条件的严重影响[46,47]。单独使用NPP-VIIRS DNB可能会产生不准确的城市空间格局，尤其是在相当不同的经济条件下。关于这个问题，郭等人[31]建立了结合NPP-VIIRS-DNB和MODIS-NDVI变量的大型不透水表面指数（LISI），大大提高了萃取性能。基于LISI的方法提供了准确的空间模式，从核心城市地区的高值到农村地区的低值，总的均方根误差为0.11。同样，Sharma等人[48]提出了一个结合modis多光谱图像和npp-viirs dnb的城市建成指数（ubi），以生成全球范围内500米分辨率的城市建成区地图。光谱数据和VIIRS夜间光照数据的结合显著降低了误分类误差，捕捉了更详细的城市空间格局信息。然而，指数只是估计线性模型的性能。为了探索非参数化算法，扩大多源粗图像在城市土地提取中的联合应用，还需要进行更多的研究[49]。

The main purpose of this study is to evaluate the effectiveness of combining NPP-VIIRS DNB and MODIS NDVI on regional urban extent mapping by using a Back Propagation (BP) neural network-based one-class classification method. To achieve this goal, we extracted the urban land area in mainland China (excluding surrounding islands) in 2012 and conducted contrastive analyses on the input data, the classification method and the results. We comparatively analyze the extracted urban map with the map generated by DMSP/OLS NLD and MODIS NDVI and the urban extent extracted from MCD12Q1 in MODIS products in Section 5.1. We explore the advantages of combining NPP-VIIRS DNB and MODIS NDVI in Section 5.2 and compare our method with the linear method (LISI) in Section 5.3.

本研究的主要目的是利用基于反向传播（BP）神经网络的一类分类方法，评估NPP-VIIRS-DNB与MODIS-NDVI相结合在区域城市范围图绘制中的有效性。为了实现这一目标，我们在2012年提取了中国大陆（不包括周边岛屿）的城市土地面积，并对输入数据、分类方法和结果进行了对比分析。我们将提取的城市地图与DMSP/OLS NLD和MODIS NDVI生成的地图以及第5.1节中MODIS产品中MCD12Q1提取的城市范围进行了对比分析。我们在第5.2节中探讨了NPP-VIIRS DNB和MODIS NDVI相结合的优点，并将我们的方法与第5.3节中的线性方法（LISI）进行了比较。

2. Materials
2.1. Defining Urban Extent
Regarding different research perspectives, different definitions of “urban extent” are introduced [50–52]. For example, if the urban studies are related to census data, “urban extent” may be defined by population distribution; if the studies use nighttime light images, the definition may
be related to economic condition. To reduce the uncertainty of the term “urban extent”, it is necessary to define it before characterizing urbanized areas.

针对不同的研究视角，介绍了“城市范围”的不同定义[50-52]。例如，如果城市研究与人口普查数据有关，“城市范围”可由人口分布定义；如果研究使用夜间光照图像，则定义可能与经济状况有关。为了减少“城市范围”这一术语的不确定性，在描述城市化区域之前必须对其进行定义。
As this study refers to the physical attributes of urban extent, we employ the definition proposed by Schneider et al. [12]; that is, “urban extent” refers to places dominated by artificial surfaces with a minimum area of 1 km2, including all non-vegetative, human-constructed elements, such as
roads, buildings and runways, and “dominated” implies coverage greater than 50% of a pixel. When vegetation covers most of a pixel, it will not be considered as the urban type, even if it may function as urban space, such as a park.

由于本研究涉及城市范围的物理属性，我们采用了Schneider等人提出的定义。[12]；也就是说，“城市范围”是指以人工表面为主，最小面积为1平方公里的区域，包括所有非植物性的、人类建造的元素，例如
道路、建筑物和跑道以及“主导”意味着覆盖率超过像素的50%。当植被覆盖了一个像素的大部分，它将不会被视为城市类型，即使它可以作为城市空间，如公园。

2.2. Study Area
This paper takes mainland China as the case study area, which contains 30 provinces, but excludes the surrounding islands (Figure 1).

本文以中国大陆为例，包括30个省，但不包括周边岛屿（图1）。

Due to the rapid economic and social development, China’s urbanization rate grew rapidly over the past approximately 38 years. Until the end of 2012, more than 0.71 billion Chinese (about 52% of the population) lived in cities [53]. However, the western economy is obviously lagging behind the eastern, and the city-dwelling populations mainly gather in the east. The economic development in China is extremely imbalanced. Besides, China has wide land coverage (approximately 9.6 million km2), and its geographical features vary greatly from region to region. In the southeast coast, the climate is a continental monsoon climate, while it changes into typical inland dry weather in northwest China. Multivariate climate conditions bring out different vegetation characteristics of cities. As annual average precipitation decreases from east to west, vegetation coverage reduces with the same change tendency. Multiple states of the economy and the geographical differences make China an ideal experimental region for estimating the combination of NPP-VIIRS DNB and MODIS NDVI at continental scales.

由于经济社会的快速发展，近38年来，我国城镇化率快速增长。到2012年底，超过7.1亿中国人（约占总人口的52%）居住在城市中[53]。然而，西部经济明显落后于东部，城市居民主要集中在东部。中国经济发展极不平衡。此外，中国土地覆盖范围广（约960万平方公里），各地的地理特征差异很大。东南沿海为大陆性季风气候，西北地区为典型的内陆干旱气候。多元气候条件下城市植被特征不同。随着年平均降水量由东向西递减，植被覆盖度也呈同样的变化趋势。经济的多个状态和地理差异使我国成为大陆尺度上核电站-VIIRS-DNB和MODIS-NDVI组合的理想试验区。

In addition, eight cities (Urumchi, Suihua, Hohhot, Datong, Beijing, Chengdu, Wuhan and Guangzhou) with different levels of urbanization and different geographical locations are selected as the cases for contrastive analysis in Section

并选取乌鲁木齐、绥化、呼和浩特、大同、北京、成都、武汉、广州等8个城市，以不同城市化水平和地理位置为例进行分区对比分析。

2.3. Data Acquisition and Pre-Processing

Four different sources of data are used for this work: (1) collection 5 MODIS 500 m NDVI data products (MOD13A1) for 2012 [54]; (2) NPP-VIIRS DNB composited data for 2012 [55]; (3) DMSP/OLS NLD acquired by the DMSP satellite F18 for 2012 [56]; and (4) Global Land Cover Mapping at 30 m resolution for 2010 (GLC30-2010) [57]. A brief description of the datasets is listed in Table 1. The input data for classification are the maximum NDVI retrieved from MODIS NDVI products, and the normalized NPP-VIIRS DNB. DMSP/OLS NLD and GLC30-2010 are auxiliary DMSP/OLS NLD is employed for filtering NPP-VIIRS DNB, and the GLC30-2010 dataset is used for sample selection. All classification data are reprojected to Albers conical equal area projection by the nearest-neighbor resampling algorithm and resampled into the same spatial resolution of 480 m. Details of the data pre-processing are described as follows.

这项工作使用了四种不同的数据来源：（1）收集5个modis 500 m ndvi数据产品（mod13a1）用于2012年[54]；（2）NPP-VIIRS dnb合成数据用于2012年[55]；（3）DMSP/OLS nld由DMSP卫星F18于2012年[56]；以及（4）2010年分辨率为30 m的全球土地覆盖图（glc30-2010）[57]。表1列出了数据集的简要描述。用于分类的输入数据是从modis ndvi产品中检索到的最大ndvi，以及标准化的npp-viirs dnb。DMSP/OLS NLD和GLC30-2010是辅助的，DMSP/OLS NLD用于过滤NPP-VIIRS DNB，GLC30-2010数据集用于样本选择。所有分类数据通过最近邻重采样算法重新投影到阿尔伯斯圆锥等面积投影中，再采样到480m的同一空间分辨率，数据预处理的细节如下。

https://www.ctahr.hawaii.edu/grem/mod13ug/sect0005.html  https://lpdaac.usgs.gov/node/838