This record is maintained as part of the global dem collection. Read the documentation before using the files in a new analysis.这条记录属于全球 DEM数据集。将文件用于新的分析前，请先阅读相关文档。
Dataset overview
GDEM_2022 is a global 3 arc-second digital elevation model at approximately 90 m resolution. It combines a global low-resolution elevation and bathymetry reference with higher-resolution land and ocean samples to reconstruct a continuous raster surface for Earth observation and terrain analysis.
The dataset addresses the resolution gap between the 15 arc-second GEBCO_2021 grid and regional high-resolution DEM products. The paper describes GDEM_2022 as a global land and ocean product, with the ocean surface inherited from GEBCO_2021 and enhanced through regional high-resolution ocean DEM samples.
Key dataset facts:
- 15 arc-second inputs are super-resolved to a 3 arc-second target at a scale factor of 5
- Approximately 579,900 globally distributed land sample units are used for training and validation
- Approximately 28,500 ocean samples support transfer learning and fine-tuning
- The generated global raster contains approximately 93.3 billion elevation pixels
- The paper reports a 23.75% average RMSE improvement over bicubic interpolation in land testing
Method
The workflow is built around DEM-SRNet, a deep residual super-resolution network derived from EDSR. The land model is pre-trained with GEBCO_2021 and NASADEM samples, then adapted to ocean regions with limited high-resolution regional DEMs.
During preprocessing, the source DEMs are converted to a consistent WGS84 and EGM96 reference, normalized to values between 0 and 1, and divided into paired patches. High-resolution NASADEM and regional ocean DEMs use 160 x 160 pixel patches, while the 15 arc-second GEBCO_2021 inputs use 32 x 32 pixel patches. Missing elevation values are removed before patch construction.
The network uses 42 residual blocks, an interpolation layer with scale factor 5, 88 convolutional layers, 43 element-wise addition layers and 50,734,080 parameters. Adam optimization starts from a learning rate of 0.0001 for pre-training and is reduced to 0.00001 during fine-tuning. Dynamic weighted loss and early stopping with a patience of 6 are used to stabilize training and reduce overfitting.
Paper-reported validation
In ocean comparisons against regional high-resolution DEM references, the paper reports a mean absolute vertical error of 19.58 m, with 63.99% of absolute errors below 10 m. The comparison with GEBCO_2021 shows increased spatial detail in both terrestrial and ocean areas.
Output
The dataset is provided as a GeoTIFF raster in geographic coordinates. It is intended for global terrain interpretation, model input, regional comparison and further geospatial analysis.
数据集概览
GDEM_2022 是一套约 90 米分辨率的全球 3 弧秒数字高程模型。它将全球低分辨率高程与水深参考数据和更高分辨率的陆地与海洋样本结合起来 重建面向地球观测与地形分析的连续栅格表面
这个数据集弥合了 15 弧秒 GEBCO_2021 网格与区域高分辨率 DEM 产品之间的分辨率差异。论文将 GDEM_2022 描述为覆盖陆地与海洋的全球产品 海洋表面继承 GEBCO_2021 并通过区域高分辨率海洋 DEM 样本进行增强
数据集信息：
- 将 15 弧秒输入以 5 倍尺度超分辨率重建为 3 弧秒目标
- 训练与验证使用约 579900 个全球分布的陆地样本单元
- 迁移学习与微调使用约 28500 个海洋样本
- 全球栅格约包含 933 亿个高程像元
- 论文报告陆地区域测试相对于双三次插值平均 RMSE 提升 23.75%
方法
整个流程以 DEM-SRNet 深度残差超分辨率网络为核心 网络结构源自 EDSR。模型先使用 GEBCO_2021 与 NASADEM 样本进行陆地预训练 再使用有限的区域高分辨率海洋 DEM 对海洋区域进行适配
预处理阶段将不同来源的 DEM 转换到一致的 WGS84 与 EGM96 参考系 将数值归一化到 0 到 1 并切分为配对样本。高分辨率 NASADEM 与区域海洋 DEM 使用 160 x 160 像元块 15 弧秒 GEBCO_2021 输入使用 32 x 32 像元块 切块前移除缺失高程值
网络包含 42 个残差块 一个 5 倍插值层 88 个卷积层 43 个逐元素加法层以及 50734080 个参数。预训练阶段 Adam 初始学习率为 0.0001 微调阶段调整为 0.00001 同时使用动态加权损失与耐心值为 6 的提前停止策略来稳定训练并减少过拟合
论文验证结果
在与区域高分辨率海洋 DEM 参考数据的对比中 论文报告平均绝对高程误差为 19.58 米 其中 63.99% 的绝对误差小于 10 米。与 GEBCO_2021 对比可以看到陆地与海洋区域的空间细节均有所增加
输出
数据集以地理坐标下的 GeoTIFF 栅格提供 可用于全球地形解读 模型输入 区域对比和后续地学分析
04 / ACCESS REQUEST04 / 数据申请
Request the data.申请数据
Tell us who you are and how you plan to use this record. Every file in this dataset will be included in the request.请告诉我们你的身份以及使用这条记录的计划。本数据集中的全部文件都会包含在申请中。