siyux1927/imaestro-sar-forest

该数据集包含处理的Sentinel-1合成孔径雷达（SAR）图像和从I-MAESTRO合成森林清单派生的森林结构属性，覆盖三个欧洲地点（法国Bauges、波兰Milicz、斯洛文尼亚Sneznik）。它专为多任务学习实验设计，用于联合森林树种分割、冠层高度回归和地上生物量回归。数据来源包括：森林清单来自I-MAESTRO合成数据集，覆盖约100,000公顷，包含42M+树木，涉及51个物种，网格分辨率为25米；SAR图像来自Sentinel-1 GRD双极化场景，采集于2019年生长季节，经过辐射校准、地形校正、重投影和重采样处理；生物量目标通过物种特异性异速生长方程从树木级数据估算，冠层高度使用每个单元格的95th百分位高度，优势树种由每个单元格的树木计数确定。数据集结构包括训练补丁（形状为(N, 5, 64, 64)的数组，包含VH和VV后向散射、生物量、优势树种和冠层高度通道）、栅格文件和辅助文件。数据集用于机器学习模型训练和评估，支持森林遥感研究。This dataset contains processed Sentinel-1 Synthetic Aperture Radar (SAR) images and forest structural attributes derived from the I-MAESTRO synthetic forest inventory, covering three European sites: Bauges (France), Milicz (Poland), and Sneznik (Slovenia). It is specifically designed for multi-task learning experiments, enabling joint forest tree species segmentation, canopy height regression, and above-ground biomass regression. Data sources are as follows: The forest inventory originates from the I-MAESTRO synthetic dataset, covering approximately 100,000 hectares, containing more than 42 million trees across 51 species, with a grid resolution of 25 meters; SAR images are sourced from Sentinel-1 GRD dual-polarization scenes acquired during the 2019 growing season, which have undergone radiometric calibration, terrain correction, reprojection, and resampling processing. Biomass targets are estimated from tree-level data using species-specific allometric equations, canopy height is derived as the 95th percentile height of each grid cell, and the dominant tree species is determined by the tree count per grid cell. The dataset structure includes training patches (arrays with shape (N, 5, 64, 64), containing VH and VV backscatter, biomass, dominant tree species, and canopy height channels), raster files, and auxiliary files. This dataset is utilized for machine learning model training and evaluation, and supports forest remote sensing research.