【免费】Google发布近实时的10m分辨率全球土地利用数据

10m

需积分: 0 23 浏览量 2023-06-07 13:10:04 上传评论收藏 3.09MB PDF 举报

身份认证购VIP最低享 7 折!

领优惠券(最高得80元）

资源推荐

资源详情

资源评论

Scientific DATA | (2022) 9:251 | https://doi.org/10.1038/s41597-022-01307-4

www.nature.com/scientificdata

Dynamic World, Near real-time



mapping





✉



























































































Regularly updated global land use land cover (LULC) datasets provide the basis for understanding the sta-

tus, trends, and pressures of human activity on carbon cycles, biodiversity, and other natural and anthropo-

genic processes

1–3

. Annual maps of global LULC have been developed by many groups. ese maps include

the National Aeronautics and Space Administration (NASA) MCD12Q1 500 m resolution dataset

4,5

(2001–

2018), the European Space Agency (ESA) Climate Change Initiative (CCI) 300 m dataset

(1992–2018), and

Copernicus Global Land Service (CGLS) Land Cover 100 m dataset

7,8

(2015–2019). While widely used, many

important LULC change processes are dicult or impossible to observe at a spatial resolution greater than

100 m and annual temporal resolution

, such as emerging settlements and small-scale agriculture (prevalent in

the developing world) and early stages of deforestation and wetland/grassland conversion. Inability to resolve

these processes introduces signicant errors in our understanding of ecological dynamics and carbon budgets.

us, there is a critical need for spatially explicit, moderate resolution (10–30 m/pixel) LULC products that are

updated with greater temporal frequency.

Currently, almost all moderate resolution LULC products are available with only limited spatial and/or tem

poral coverage (e.g., USGS NLCD

and LCMAP

) or via proprietary and/or closed products (e.g., BaseVue

GlobeLand30

, GlobeLand10

) that are generally not available to support monitoring, forecasting, and decision

making in the public sphere. A noteworthy exception is the recent iMap 1.0

series of products available globally

at a seasonal cadence with a 30 m resolution. Nonetheless, globally consistent, near real-time (NRT) mapping

of LULC remains an ongoing challenge due to the tremendous computational and data storage requirements.

Simultaneous advances in large-scale cloud computing and machine learning algorithms in

high-performance open source soware frameworks (e.g., TensorFlow

) as well as increased access to satellite

Google, LLC, 1600 Amphitheatre Pkwy., Mountain View, CA, 94043, USA.

National Geographic Society, 1145 17th

St NW, Washington, DC, 20036, USA.

World Resources Institute, 10 G St NE #800, Washington, DC, 20002, USA.

Department of Earth & Environment, Boston University, 685 Commonwealth Avenue, Boston, MA, 02215, USA.

✉

e-mail: c@google.com





Scientific DATA | (2022) 9:251 | https://doi.org/10.1038/s41597-022-01307-4

www.nature.com/scientificdata

www.nature.com/scientificdata/

image collections through platforms such as Google Earth Engine

have opened new opportunities to create

global LULC datasets at higher spatial resolutions and greater temporal cadence than ever before. In this paper,

we introduce a new NRT LULC dataset produced using a deep-learning modeling approach. Our model, which

was trained using a combination of hand-annotated imagery and unsupervised methods, is used to operation

ally generate NRT predictions of LULC class probabilities for new and historic Sentinel-2 imagery using cloud

computing on Earth Engine and Google Cloud AI Platform. ese products, which we refer to collectively as

Dynamic World, are available as a continuously updating Earth Engine Image Collection that enables users to

leverage both class probabilities and multi-temporal results to track LULC dynamics in NRT and create custom

products suited to their specic needs. We nd that our model exhibits strong agreement with expert annota

tions for an unseen validation dataset, and though dicult to compare with existing products due to dierences

in temporal resolution and classication schemes, achieves better or comparable performance relative to other

state-of-the-art global and regional products when compared to the same reference dataset.



 e classication schema or “taxonomy” for Dynamic World, shown

in Table1, was determined aer a review of global LULC maps, including the USGS Anderson classication

system

, ESA Land Use and Coverage Area frame Survey (LUCAS) land cover modalities

, MapBiomas classi-

cation

, and GlobeLand30 land cover types

. e Dynamic World taxonomy maintains a close semblance to

the land use classes presented in the IPCC Good Practice Guidance (forest land, grassland, cropland, wetland,

settlement, and other)

to ensure easier application of the resulting data for estimating carbon stocks and green-

house gas emissions. Unlike single-pixel labels, which are usually dened in terms of percent cover thresholds,

the Dynamic World taxonomy was applied using “dense” polygon-based annotations such that LULC labels are

applied to areas of relatively homogenous cover types with similar colors and textures.

 Our modeling approach relies on semi-supervised deep learning and requires

spatially dense (i.e., ideally wall-to-wall) annotations. To collect a diverse set of training and evaluation data, we

divided the world into three regions: the Western Hemisphere (160°W to 20°W), Eastern Hemisphere-1 (20°W

to 100°E), and Eastern Hemisphere-2 (100°E to 160°W). We further divided each region by the 14 RESOLVE

Ecoregions biomes

. We collected a stratified sample of sites for each biome per region based on NASA

MCD12Q1 land cover for 2017

. Given the availability of higher-resolution LULC maps in the United States and

Brazil, we used the NLCD 2016

and MapBiomas 2017

LULC products respectively in place of MODIS prod-

ucts for stratication in these two countries.

At each sample location, we performed an initial selection of Sentinel-2 images from 2019 scenes based on

image cloudiness metadata reported in the Sentinel-2 tile’s QA60 band. We further ltered scenes to remove

images with many masked pixels. We nally extracted individual tiles of 510 × 510 pixels centered on the sample

sites from random dates in 2019. Tiles were sampled in the UTM projection of the source image and we selected

one tile corresponding to a single Sentinel-2 ID number and single date.

Further steps were then taken to obtain an “as balanced as possible” training dataset with respect to the

LULC classications from the respective LULC products. In particular, for each Dynamic World LULC category

contained within a tile, the tile was labeled to be high, medium, or low in that category. We then selected an

approximately equal number of tiles with high, medium or low category labels for each category.

To achieve a large dataset of labeled Sentinel-2 scenes, we worked with two groups of annotators. e rst

group included 25 annotators with previous photo-interpretation and/or remote sensing experience. e expert

group labeled approximately 4,000 image tiles (Fig.1a), which were then used to train and measure the per

formance and accuracy of a second “non-expert” group of 45 additional annotators who labeled a second set

of approximately 20,000 image tiles (Fig.1b). A nal validation set of 409 image tiles were held back from

the modeling eort and used for evaluation as described in the Technical Validation section. Each image tile

in the validation set was annotated by three experts and one non-expert to facilitate cross-expert and expert/

non-expert QA comparisons.

All Dynamic World annotators used the Labelbox platform

, which provides a vector drawing tool to

mark the boundaries of feature classes directly over the Sentinel-2 tile (Fig.2). We instructed both expert and

non-expert annotators to use dense markup instead of single pixel labels with a minimum mapping unit of 50

× 50 m (5 × 5 pixels). For water, trees, crops, built area, bare ground, snow & ice, and cloud, this was a fairly

straightforward procedure at the Sentinel-2 10 m resolution since these feature classes tend to appear in fairly

homogenous agglomerations. Shrub & scrub and ooded vegetation classes proved to be more challenging as

they tended not to appear as homogenous features (e.g. mix of vegetation types) and have variable appearance.

Annotators used their best discretion in these situations based on the guidance provided in our training material

(i.e. descriptions and examples in Table1). In addition to the Sentinel-2 tile, annotators had access to a match

ing high-resolution satellite image via Google Maps and ground photography via Google Street View from the

image center point. We also provided the date and center point coordinates for each annotation. All annotators

were asked to label at least 70% of a tile within 20 to 60 minutes and were allowed to skip some tiles to best bal

ance their labeling accuracy with their eciency.

 We prepared Sentinel-2 imagery in a number of ways to accommodate both annota-

tion and training workows. An overview of the preprocessing workow is shown in Fig.3.

For training data collection, we used the Sentinel-2 Level-2A (L2A) product, which provides radiometri-

cally calibrated surface reectance (SR) processed using the Sen2Cor soware package

. is advanced level

of processing was advantageous for annotation, as it attempts to remove inter-scene variability due to solar dis

tance, zenith angle, and atmospheric conditions. However, systematically produced Sentinel-2 SR products are

Scientific DATA | (2022) 9:251 | https://doi.org/10.1038/s41597-022-01307-4

www.nature.com/scientificdata

www.nature.com/scientificdata/

currently only available from 2017 onwards. erefore, for our modeling approach, we used the Level-1C (L1C)

product, which has been generated since the beginning of the Sentinel-2 program in 2015. e L1C product rep

resents Top-of-Atmosphere (TOA) reectance measurements and is not subject to a change in processing algo-

rithm in the future. We note that for any L2A image, there is a corresponding L1C image, allowing us to directly

map annotations performed using L2A imagery to the L1C imagery used in model training. All bands except

for B1, B8A, B9, and B10 were kept, with all bands bilinearly upsampled to 10 m for both training and inference.

In addition to our preliminary cloud ltering in training image selection, we adopted and applied a novel

masking solution that combines several existing products and techniques. Our procedure is to rst take the

10 m Sentinel-2 Cloud Probability (S2C) product available in Earth Engine

and join it to our working set of

Sentinel-2 scenes such that each image is paired with the corresponding mask. We compute a cloud mask by

thresholding S2C using a cloud probability of 65% to identify pixels that are likely obscured by cloud cover. We

then apply the Cloud Displacement Index (CDI) algorithm

and threshold the result to produce a second cloud

Class ID LULC Type Description Examples

0 Water

• Water is present in the image.

• Contains little-to-no sparse vegetation, no rock

outcrop, and no built-up features like docks.

• Does not include land that can or has previously been

covered by water.

• Rivers

• Ponds & Lakes

• Ocean

• Flooded Salt Pans

1 Trees

• Any signicant clustering of dense vegetation,

typically with a closed or dense canopy.

• Taller and darker than surrounding vegetation (if

surrounded by other vegetation).

• Wooded vegetation

• Dense green shrubs

• Cluster of dense, tall vegetation within savannas

• Plantations such as apples, bananas, citrus, and

rubber

• Swamp (dense/tall vegetation with no obvious

water)

• Any mix of the above

• Any burned areas of the above

2 Grass

• Open areas covered in homogenous grasses with little

to no taller vegetation.

• Other homogenous areas of grass-like vegetation

(blade-type leaves) that appear dierent from trees

and shrubland.

• Wild cereals and grasses with no obvious human

plotting (i.e. not a structured eld).

• Natural meadows and elds with sparse or no

tree cover

• Open savanna with little to no tree cover

• Parks, golf courses, human manicured lawns,

including large elds in urban settings like

soccer and baseball.

• Tree cut-throughs for power lines, gas etc.

• Pastures

• Reeds and marshes with no obvious ooding

3 Flooded vegetation

• Areas of any type of vegetation with obvious

intermixing of water.

• Do not assume an area is ooded if ooding is

observed in another image.

• Seasonally ooded areas that are a mix of grass/

shrub/trees/bare ground.

• Flooded mangroves

• Emergent vegetation

4 Crops • Human planted/plotted cereals, grasses, and crops.

• Corn, wheat, soy, etc.

• Hay and fallow plots of structured land

5 Shrub & Scrub

• Mix of small clusters of plants or individual plants

dispersed on a landscape that shows exposed soil

and rock.

• Scrub-lled clearings within dense forests that are

clearly not taller than trees. Appear grayer/browner

due to less dense leaf cover.

• Moderate to sparse cover of bushes, shrubs, and

tus of grass

• Savannas with very sparse grasses, trees, or other

plants

6 Built area

• Clusters of human-made structures or individual

very large human-made structures.

• Contained industrial, commercial, and private

building, and the associated parking lots.

• A mixture of residential buildings, streets, lawns,

trees, isolated residential structures or buildings

surrounded by vegetative land covers.

• Major road and rail networks outside of the

predominant residential areas.

• Large homogeneous impervious surfaces, including

parking structures, large oce buildings, and

residential housing developments containing clusters

of cul-de-sacs.

• Cluster of houses, can include smalls lawns or

small patches of trees can be included

• Dense villages, town, and cityscape (buildings

and roads together)

• Clusters of paved roads and large highways

• Asphalt and other human-made surfaces

7 Bare ground

• Areas of rock or soil containing very sparse to no

vegetation.

• Large areas of sand and deserts with no to little

vegetation.

• Large individual or dense networks of dirt roads.

• Exposed rock

• Exposed soil

• Desert and sand dunes

• Dry salt ats and salt pans

• Dried lake bottoms

• Mines

• Large empty lots in urban areas

8 Snow & Ice

• Large homogenous areas of thick snow or ice,

typically only in mountain areas or highest latitudes.

• Large homogenous areas of snowfall.

• Glaciers

• Permanent snowpack

• Snowfall

Table 1. Dynamic World Land Use Land Cover (LULC) classication taxonomy. Denitions and examples

were provided as part of annotator reference materials, along with descriptions of colors and patterns typically

associated with each LULC type.

Scientific DATA | (2022) 9:251 | https://doi.org/10.1038/s41597-022-01307-4

www.nature.com/scientificdata

www.nature.com/scientificdata/

mask, which is intersected with the S2C mask to reduce errors of commission by removing bright non-cloud

targets based on Sentinel-2 parallax eects. We nally intersect this sub-cirrus mask with a threshold on the

Sentinel-2 cirrus band (B10) using the thresholding constants proposed for the CDI algorithm

, and take a

morphological opening of this as our cloudy pixel mask. is mask is computed at 20 m resolution.

In order to remove cloud shadows, we extend the cloudy pixel mask 5 km in the direction opposite the solar

azimuthal angle using the scene level metadata “SOLAR_AZIMUTH_ANGLE” and a directionaldistancetrans

form (DDT) operation in Earth Engine. e nal cloud and shadow mask is resampled to 100 m to decrease both

the data volume and processing time. e resulting mask is applied to Sentinel-2 images used for training and

inference such that unmasked pixels represent observations that are likely to be cloud- and shadow-free.

e distribution of Sentinel-2 reectance values are highly compressed towards the low end of the sensor

range, with the remainder mostly occupied by high return phenomena like snow and ice, bare ground, and

specular reection. To combat this imbalance, we introduce a normalization scheme that better utilizes the

useful range of Sentinel-2 reectance values for each band. We rst log-transform the raw reectance values to

Fig. 1 Global distribution of annotated Sentinel-2 image tiles used for model training and periodic testing

(neither including 409 validation tiles). (a) 4,000 tiles interpreted by a group of 25 experts (b) 20,000 tiles

interpreted by a group of 45 non-experts. Hexagons represent approximately 58,500 km

areas and shading

corresponds to the count of annotated tile centroids per hexagon.

Fig. 2 Sentinel-2 tile and example reference annotation provided as part of interpreter training. is example

was used to illustrate the Flooded vegetation class, which is distinguished by small “mottled” areas of water

mixed with vegetation near a riverbed. Also note that some areas of the tile are le unlabeled.

剩余16页未读，继续阅读

评论收藏

内容反馈

NagaClimber

粉丝: 1
资源: 1

Google发布近实时的10m分辨率全球土地利用数据

论文研究 - 基于多时相遥感数据的多变量变化矢量分析（MCVA）估算坦莱铝土矿的土地利用变化

谷歌2016图像超分辨率论文源码

Google全球数据中心与谷歌中国机房

基于Google Earth Engine与机器学习的大尺度30m分辨率沙地灌木覆盖度估算.pdf

Google 全球公交数据标准GTFS样例数据

谷歌卫星地图不同级别对应的比例尺、分辨率、视点高度之间的关系

Google数据仓库Mesa论文

Chrome分辨率插件

利用GOOGLE EARTH 提取海拔数据

googlefinance, python 模块从Google获取实时股票数据.zip

NWPU VHR-10数据集.rar

谷歌地图全球版谷歌地图全球版

Google Play 商店 10 年数据纵览

谷歌高清地图下载软件破解版

GoogleEarth谷歌地球完整离线安装包

全球采矿区域边界数据集（1km）-v2-30arcsecond.tif

谷歌大名鼎鼎的RAISR超分辨率论文

GoogleEarth高程数据采集工具

SIRI-WHU Data Set 遥感影像数据集.7z

工训垃圾分类的数据集的

2023年中国研究生数学建模竞赛思路

免费chatGPT4模拟版下载安装可直接使用

最新版YS9082HC主控开卡工具 YS9082HC-MPToolV8.00.00.18.826-HCS1A25E2023062

数学建模2022国赛C题省一论文

openai/chatgpt微调/fine-tuning/测试用/投喂资源

IEEE 802.11be（WiFi7） 协议原文pdf文档

青霉素发酵过程仿真数据

数据集图片爬取工具image_tool

2023研究生数学建模F题强对流降水临近预报，全套详细代码加思路，九大文件手把手教你，附有结果图

最新资源

IEEE 802.11be（WiFi7）协议原文pdf文档