This record is maintained as part of the australian coastal dem collection. Read the documentation before using the files in a new analysis.这条记录属于澳大利亚沿海 DEM数据集。将文件用于新的分析前，请先阅读相关文档。
Dataset overview
AUSDEM_2023 is an open-access 1 arc-second Australian coastal digital elevation model. It reconstructs a high-resolution land and ocean surface from the 15 arc-second GEBCO_2023 grid, using NASADEM as the 1 arc-second land reference and high-resolution AusSeabed bathymetry as the ocean reference.
The product covers the Australian coastal domain shown in the paper, approximately 0-50 S and 100-160 E, and is distributed as 5 x 5 degree tiles. It combines real low- and high-resolution measurements rather than creating the low-resolution input by manually downsampling the target image.
- 15 arc-second GEBCO_2023 input is reconstructed to 1 arc-second output with a cumulative 15x scale factor
- The 1 arc-second output corresponds to approximately 30 m at the Australian latitude range
- The transfer-learning and test areas cover approximately 300,000 square kilometers of ocean around Australia
- The study constructed 22,725 paired sample groups and used an 80/20 training-validation split
- Land pre-training used 909 NASADEM samples; AusSeabed ocean samples were used for transfer learning and independent testing
- The paper reports an average RMSE reduction of 11.64% and a PSNR increase of 1.93% over bicubic interpolation
Data sources and preprocessing
| Dataset | Role | Resolution | Coverage and use |
|---|---|---|---|
| NASADEM | High-resolution land reference | 1 arc-second | Australian land samples for pre-training and validation |
| GEBCO_2023 | Low-resolution land and ocean input | 15 arc-second | Global grid used to generate the low-resolution paired inputs |
| AusSeabed Marine Data Portal | High-resolution ocean reference | 30 m | Northern Australia and Great Barrier Reef bathymetry for transfer learning and testing |
All sources are projected to WGS84 and adjusted to the EGM96 geoid reference. Each sample is normalized independently using the paper’s range, with Imin = -10956 and Imax = 8627. Missing or invalid values are discarded before patch construction. High-resolution DEMs are split into 480 x 480 pixel patches and low-resolution DEMs into 32 x 32 pixel patches. Training samples are augmented with 90, 180 and 270 degree rotations, increasing the orientation variants fourfold.
The preprocessing workflow uses ArcGIS 10.3 and Python GDAL, with the prepared pairs passed to a PyTorch training pipeline. This keeps the horizontal and vertical references consistent while preserving the spatial structure needed for coastal and bathymetric reconstruction.
Method
AUSDEM-SRNet is a deep gradient prior network designed for the scarcity of high-resolution ocean DEM samples. The model is first pre-trained on the larger land domain using paired GEBCO_2023 and NASADEM data. The layers before the upsampling modules are then copied and frozen for ocean transfer learning, while the target-side layers are fine-tuned with GEBCO_2023 ocean input and AusSeabed high-resolution bathymetry.
The network contains 80 residual blocks, 64 feature maps, 166 convolutional layers, 80 activation layers, 82 element-wise addition layers and 2 sub-pixel upsampling modules. It uses residual scaling beta = 0.1 and LReLU alpha = 0.01 in the architecture described by the paper, with 16,959,744 trainable parameters. The two sub-pixel stages scale the input by x5 and x3, giving a cumulative 15x reconstruction from 15 arc-seconds to 1 arc-second.
Training uses an L1 pixel loss and the Adam optimizer with beta1 = 0.9, beta2 = 0.999. The initial learning rate is 0.0001 for land pre-training and 0.00001 for ocean fine-tuning, with a batch size of 32, early stopping patience of 6 and about 1000 epochs typically sufficient for training.
Validation and results
The paper evaluates RMSE and PSNR against measured high-resolution DEM samples. It compares bicubic interpolation, SRCNN, WDSR, EDSR, an AUSDEM-SRNet land pre-training model, an ocean pre-training model and the final transfer-learning model across ten held-out regions.
| Model | Average RMSE change vs bicubic | Average PSNR change vs bicubic |
|---|---|---|
| Aus-land pre-training | -9.00% | +1.42% |
| Aus-ocean pre-training | -9.28% | +1.54% |
| AUSDEM-SRNet transfer learning | -11.64% | +1.93% |
The visual comparisons show that the final model preserves sharper coastline boundaries, island contours and high-frequency bathymetric texture than bicubic interpolation and SRCNN. WDSR and EDSR improve detail recovery, while transfer learning from high-resolution ocean DEMs gives AUSDEM-SRNet the strongest reconstruction across the test regions. The error maps also show a lower error distribution for the final model, while residual striped patterns can still reflect systematic artifacts in the GEBCO_2023 low-resolution input.
Output and access
The final product is an open-access GeoTIFF DEM split into 5 x 5 degree tiles. Dataset files are supplied through the access request form on this page. The paper, model implementation and training resources are recorded in the project documentation for reference.
数据集概览
AUSDEM_2023 是一套开放获取的澳大利亚沿海 1 弧秒数字高程模型。它以 15 弧秒 GEBCO_2023 为低分辨率输入，使用 NASADEM 作为 1 弧秒陆地参考，并使用 AusSeabed 高分辨率测深数据作为海洋参考，通过深度学习重建连续的陆地与海洋高程表面。
数据集覆盖论文所示的澳大利亚沿海区域，大致范围为 0-50 S、100-160 E，并以 5 x 5 度分块提供。论文使用真实的低分辨率与高分辨率测量数据配对，而不是先将目标高分辨率图像人工降采样来制造输入。
- 将 15 弧秒 GEBCO_2023 输入以累计 15 倍尺度重建为 1 弧秒输出
- 1 弧秒输出在澳大利亚纬度范围内约对应 30 米空间尺度
- 海洋迁移学习与测试区域覆盖澳大利亚周边约 30 万平方千米海域
- 构建约 22725 组配对样本，并按 80% / 20% 划分训练集与验证集
- 使用 909 个 NASADEM 陆地样本进行预训练，使用 AusSeabed 海洋样本进行迁移学习与独立测试
- 相比双三次插值，论文报告平均 RMSE 降低 11.64%，PSNR 提升 1.93%
数据来源与预处理
| 数据集 | 用途 | 分辨率 | 覆盖范围与作用 |
|---|---|---|---|
| NASADEM | 高分辨率陆地参考 | 1 弧秒 | 澳大利亚陆地样本，用于预训练与验证 |
| GEBCO_2023 | 低分辨率陆地与海洋输入 | 15 弧秒 | 生成配对低分辨率输入的全球网格 |
| AusSeabed Marine Data Portal | 高分辨率海洋参考 | 30 米 | 北澳大利亚与大堡礁测深数据，用于迁移学习与测试 |
所有来源首先统一到 WGS84，并调整到 EGM96 大地水准面参考。每个样本独立归一化，论文使用 Imin = -10956 与 Imax = 8627 作为值域范围。切块前删除缺失或无效值。高分辨率 DEM 切分为 480 x 480 像元块，低分辨率 DEM 切分为 32 x 32 像元块。训练样本使用 90、180 和 270 度旋转增强，使方向变体增加为原来的四倍。
预处理使用 ArcGIS 10.3 与 Python GDAL，生成的配对样本进入 PyTorch 训练流程。这样既统一了水平与垂直参考，也保留了沿海地形与海底地形重建所需的空间结构。
方法
AUSDEM-SRNet 是针对高分辨率海洋 DEM 样本稀缺问题设计的深度梯度先验网络。模型先使用较大规模的 GEBCO_2023 与 NASADEM 陆地配对数据进行预训练，再复制并冻结上采样模块之前的网络层，使用 GEBCO_2023 海洋输入与 AusSeabed 高分辨率测深数据进行海洋迁移学习。
网络包含 80 个残差块、64 个特征图、166 个卷积层、80 个激活层、82 个逐元素加法层和 2 个子像素上采样模块，共有 16959744 个可训练参数。论文结构图中使用残差缩放 beta = 0.1 与 LReLU alpha = 0.01。两个子像素阶段分别进行 x5 与 x3 放大，将 15 弧秒输入累计放大 15 倍得到 1 弧秒输出。
训练使用 L1 像元损失与 Adam 优化器，beta1 = 0.9、beta2 = 0.999。陆地预训练初始学习率为 0.0001，海洋微调学习率为 0.00001，批大小为 32，提前停止耐心值为 6，通常约 1000 个 epoch 即可完成训练。
验证与结果
论文使用 RMSE 与 PSNR 对高分辨率实测 DEM 样本进行评价，并在 10 个独立测试区域比较双三次插值、SRCNN、WDSR、EDSR、陆地预训练模型、海洋预训练模型与最终迁移学习模型。
| 模型 | 相对双三次插值的平均 RMSE 变化 | 相对双三次插值的平均 PSNR 变化 |
|---|---|---|
| 陆地预训练 | -9.00% | +1.42% |
| 海洋预训练 | -9.28% | +1.54% |
| AUSDEM-SRNet 迁移学习 | -11.64% | +1.93% |
可视化结果显示，最终模型相比双三次插值与 SRCNN 能够更好地保持海岸线边界、岛屿轮廓和高频海底纹理。WDSR 与 EDSR 对细节恢复有所改善，而高分辨率海洋 DEM 的迁移学习使 AUSDEM-SRNet 在多个测试区域获得更强的重建效果。误差图也显示最终模型的误差分布更低，但 GEBCO_2023 低分辨率输入中的系统性条纹误差仍可能传递到输出中。
输出与获取
最终产品为开放获取的 GeoTIFF DEM，按 5 x 5 度分块提供。数据文件通过本页的申请表获取。论文、模型实现与训练资源已记录在项目文档中，供研究者查阅。
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。