Tonic/landshift-sft-v1-duplicate
收藏资源简介:
LandShift SFT数据集是一个基于时间序列Sentinel-2卫星图像对的数据集,主要用于突出显示土地覆盖类型的变化。数据集通过NDVI差异来强调变化,并提供了变化描述文本和可选的连接变化区域的基础信息。数据集包含训练集(187条记录)、验证集(52条记录)和空测试集。数据集的布局包括JSONL格式的训练、验证和测试文件,图像文件(PNG格式)以及元数据文件。数据集的分割基于event_id的稳定哈希值,确保相同event_id始终位于同一分割中。数据集可以通过特定脚本重新生成并上传至Hugging Face Hub。数据集的局限性包括仅使用光学Sentinel-2数据、检测和描述文本是启发式的等。
The LandShift SFT dataset consists of temporal Sentinel-2 pairs (longer baseline by default) over sampled locations, with NDVI delta highlighting land-cover–style change. The dataset emits change captions and optional grounding for connected change regions. It includes train (187 records), validation (52 records), and an empty test set. The dataset layout comprises JSONL files for train, validation, and test, PNG image files, and metadata files. Splits are assigned based on a stable hash of event_id. The dataset can be regenerated locally using specific scripts and uploaded to the Hugging Face Hub. Limitations include being optical-only Sentinel-2, heuristic detections and captions, and conservative assistant text wording.




