Sciamlab/landslide-prevention-italy

这是一个用于预测滑坡前风险的Sentinel-2光谱复合图像的多图像视觉语言微调数据集。每个样本包含六张512×512图像（RGB + 五种光谱指数）和结构化的JSON评估，评估内容包括斜坡危险、人类影响暴露和未来最有可能发生滑坡的像素级关注区域。数据集为微调LiquidAI/LFM2.5-VL-450M模型而生成。总样本数为111，训练/测试分割为89/22，图像分辨率为512×512，地理范围为意大利地区，2025年滑坡事件。数据来源为Sentinel-2 L2A表面反射率图像和ISPRA IFFI意大利国家滑坡清单。输出标签详细描述了风险级别、土壤湿度异常、植被压力、裸露土壤暴露、水积聚、陡峭地形可见性、侵蚀证据、粘土土壤检测等。使用数据集时需要注意光谱指数是代理而非测量值，无DEM，雪/云混淆，AOI坐标与RGB图像1:1对应，类别不平衡等问题。使用数据集时需要引用ISPRA IFFI清单和Copernicus Sentinel-2任务。

Multi-image vision-language fine-tuning dataset for predicting pre-landslide risk from Sentinel-2 spectral composites. Each sample contains six 512×512 images (one RGB + five spectral indices) paired with a structured JSON assessment of slope hazard, human-impact exposure, and pixel-level areas-of-interest where future failure is most likely. Generated for fine-tuning LiquidAI/LFM2.5-VL-450M. Total samples: 111, Train / Test split: 89 / 22, Image resolution: 512 × 512, Geography: Italian regions, 2025 landslide events. Source data: Sentinel-2 L2A surface reflectance and ISPRA IFFI Italian National Landslide Inventory. Output schema includes risk level, soil moisture anomaly, vegetation stress, bare soil exposure, water accumulation, steep terrain visibility, erosion evidence, clay soil detection, etc. Considerations for using: spectral indices are proxies, not measurements; no DEM; snow/cloud confusion; AOI coordinates are 1:1 with RGB image; class imbalance. Citation required for ISPRA IFFI inventory and Copernicus Sentinel-2 mission.