GHS-BUILT-S R2023A - GHS built-up surface grid, derived from Sentinel2 composite and Landsat, multitemporal (1975-2030)

The spatial raster dataset depicts the distribution of the built-up (BU) surfaces estimates between 1975 and 2030 in 5 years intervals and two functional use components a) the total BU surface and b) the non-residential (NRES) BU surface. The data is made by spatial-temporal interpolation of five observed collections of multiple-sensor, multiple-platform satellite imageries. Landsat (MSS, TM, ETM sensor) supports the 1975, 1990, 2000, and 2014 epochs. Sentinel2 (S2) composite (GHS-composite-S2 R2020A) supports the 2018 epoch. The built-up surface fraction (BUFRAC) is estimated at 10m of spatial resolution from the S2 image data, using as learning set a composite of data from GHS-BUILT-S2 R2020A, Facebook, Microsoft, and Open Street Map (OSM) building delineation. The BUFRAC inference is made from the combination of quantized image features (reflectance, derivative of morphological profile DMP) through associative rule learning applied to spatial data analytics, which was introduced as symbolic machine learning (SML). The non-residential (NRES) domain is predicted from S2 image data by observation of radiometric, textural, and morphological features in an object-oriented image processing framework. The multi-temporal dimension is provided by testing by the SML the association between the combination of the quantized radiometric information collected by the Landsat imagery in the past epochs, and the “built-up” (BU) and “non-built-up” (NBU) class abstraction on image segments extracted from S2 images. The spatial-temporal interpolation is solved by rank-optimal spatial allocation using explanatory variables related to the landscape (slope, elevation, distance to water, and distance to vegetation) and related to the observed dynamic of BU surfaces in the past epochs.本空间栅格数据集展示了1975年至2030年间以5年为间隔的建成地表（built-up, BU）分布估算结果，包含两类功能用途组分：a) 总建成地表，b) 非住宅（non-residential, NRES）建成地表。该数据集通过对五组多传感器、多平台卫星影像观测数据进行时空插值生成。其中，Landsat（MSS、TM、ETM传感器）对应1975、1990、2000及2014年时段；Sentinel2（S2）合成数据（GHS-composite-S2 R2020A）对应2018年时段。建成地表占比（built-up surface fraction, BUFRAC）以10米空间分辨率从S2影像数据中估算得到，训练集采用GHS-BUILT-S2 R2020A、Facebook、Microsoft及开放街道地图（Open Street Map, OSM）的建筑轮廓数据合成数据集。BUFRAC的推求通过结合量化影像特征（反射率、形态剖面导数DMP）实现，采用应用于空间数据分析的关联规则学习方法，该方法以符号机器学习（symbolic machine learning, SML）形式提出。非住宅（NRES）地表组分通过面向对象影像处理框架，基于S2影像数据的辐射特征、纹理特征与形态特征观测结果进行预测。多时间维度的构建通过符号机器学习（SML）完成测试：关联过往时段Landsat影像采集的量化辐射信息组合，与从S2影像提取的影像分割块上的“建成地表（BU）”与“非建成地表（non-built-up, NBU）”类别抽象结果之间的关联。时空插值通过秩最优空间分配方法求解，所用解释变量包括与景观相关的坡度、高程、距水体距离、距植被距离，以及过往时段观测到的建成地表动态变化特征。

搜集汇总

背景与挑战

背景概述

该数据集为全球建成区表面网格数据，时间跨度为1975年至2030年，每5年间隔提供一次，空间分辨率包括10米（2018年）、100米和1公里。数据基于Sentinel2和Landsat卫星影像，通过符号机器学习方法进行时空插值生成，涵盖总建成区和非住宅建成区两个功能组件，适用于全球城市化和土地利用研究。