mpg-ranch/drone-lsr

该数据集包含三个不同时间（上午10点、中午12点和下午3点）捕获的航空正射影像图块。数据集分为三个配置：`default`（原始图像+树冠高度）、`dinov2_base`（DINOv2嵌入）和`dinov3_sat`（DINOv3嵌入）。所有配置都使用一致的训练/测试分割，并具有匹配的图块标识符以进行交叉引用。该数据集旨在训练能够保持一致特征表示的视觉编码器，即使在光照变化的情况下也是如此，适用于遥感、环境监测等领域。

This dataset contains aerial orthomosaic tiles captured at three different times of the day (10:00, 12:00, and 15:00). The dataset is organized into three configurations: `default` (raw images + canopy height), `dinov2_base` (DINOv2 embeddings), and `dinov3_sat` (DINOv3 embeddings). All configurations share consistent train/test splits with matching tile identifiers for cross-referencing. The dataset is designed for training vision encoders that maintain consistent feature representations despite changes in illumination, with applications in remote sensing and environmental monitoring.