GeoPathfinder
收藏资源简介：
GeoPathfinder是一个基于真实卫星影像的长距离空间推理基准测试数据集，灵感来源于Pathfinder任务。它旨在评估模型在真实地理环境中进行复杂空间关系判断的能力。数据样本为256×256像素的Landsat卫星影像补丁，对应地面约7.7×7.7公里范围，补丁上标有两个红点。核心任务是二分类：预测这两个点是否通过陆地相连（connect = 1）或被水体隔开（disconnect = 0）。标签来源于Overture Maps的水体多边形数据，并经过栅格化和交叉验证。数据集包含5,553个中心点互不重叠的影像补丁，覆盖全球六大洲的42个河流流域，通过生成连接和断开样本，最终得到10,546张图像，实现了精确的类别平衡。点放置采用对抗性策略，从候选位置中选出最具迷惑性的位置，以防止模型通过简单视线捷径解决问题。数据集提供固定划分和基于流域不相交的划分协议，以及简单/中等/困难点放置变体套件，用于评估泛化性能。文件包括带红点的基准测试图像、原始真彩色图像、GeoTIFF文件、掩码、叠加图像、元数据文件和水体多边形数据等。适用于图像分类任务，特别是遥感、地球观测领域的长距离空间推理基准测试。
GeoPathfinder is a long-range spatial reasoning benchmark dataset based on real satellite imagery, inspired by the Pathfinder task. It aims to evaluate models ability to make complex spatial relationship judgments in real-world geographic environments. Each data sample is a 256×256 pixel Landsat satellite image patch (native 30-meter resolution, corresponding to approximately 7.7×7.7 km on the ground) with two red points marked on it. The core task is binary classification: predicting whether the two points are connected by land (connect = 1) or separated by water bodies (disconnect = 0). Labels are derived from Overture Maps water polygon data, which is rasterized onto the exact pixel grid of each image patch and cross-validated with Landsats own QA water flags. The dataset consists of 5,553 image patches with non-overlapping centers (spaced more than one patch width apart), covering 42 river basins across six continents globally. Among these, 5,273 patches constitute the benchmark set, generating a connected sample and a distance-matched (within 25% difference) disconnected sample per patch, resulting in 10,546 images with precise class balance to ensure point distance does not carry label information. Point placement employs an adversarial strategy: each published point pair is selected from a large pool of candidate locations (96 for train/validation sets, 256 for the test set) as the most confusing positions to prevent models from solving the task via trivial line-of-sight shortcuts. The dataset provides a fixed patch-level 80/10/10 split (paired images do not cross splits) and excludes snowy scenes and unannotable patches. Additionally, it includes a basin-disjoint split protocol (splits_basin.json) specifically for testing geographic generalization. A fixed suite of easy/medium/hard point placement variants (eval_variants/) is also provided to assess model generalization across different difficulty levels (requiring reporting accuracy per level and macro-average). Files and data formats include: 1) pathfinder/connect/ and pathfinder/disconnect/ directories containing benchmark JPEG images with red points (corresponding to labels 1 and 0); 2) images/ directory with original true-color JPEG images without red points; 3) raw/ directory with original uint16 GeoTIFF files storing surface reflectance digital numbers (DN), along with scaling/offset metadata and scene source information; 4) masks/ directory with PNG format land/water/invalid/suspected area masks; 5) overlays/ directory with JPEG images for quality checks, overlaying water masks on true-color images; 6) metadata.parquet file containing metadata for each patch, such as split, region, scene ID, bounds, coordinate reference system/transform, quality check statistics, and point pixel/lat-lon coordinates; 7) water_polygons.parquet file (GeoParquet format) containing Overture water polygon geometries clipped to each patch extent; 8) multiple JSON files defining splits, excluded snowy scene IDs, and construction configurations. It is suitable for image classification tasks, particularly as a benchmark for long-range spatial reasoning in remote sensing and Earth observation, applicable for evaluating and comparing computer vision models, geospatial foundation models, and specialized spatial relationship understanding models in real, complex geographic scenarios.
数据集概述：GeoPathfinder
GeoPathfinder 是一个基于真实卫星影像的长距离空间推理基准数据集，旨在评估模型判断卫星图像中两个红点之间是否由陆地连接（connect = 1）或被水域分隔（disconnect = 0）的能力。
核心任务
- 输入：256×256 像素的 Landsat 图像（30米分辨率，覆盖约7.7×7.7公里区域），图像上标记有2个红点。
- 输出：二分类标签，表示两点之间是否为陆地连接（1）或水域阻隔（0）。
数据集规模
- 总样本量：10K < n < 100K
- 零重叠图像块：5,553个（分布在6大洲42个流域，图像块中心间距至少为一个图像块宽度）
- 基准测试集：5,273个图像块 → 共10,546张图像（完全类别平衡，每个图像块生成一个“连接”和一个“断开”样本，两点间距控制在25%范围内以消除距离偏差）
- 数据划分：固定80/10/10（训练/验证/测试）的图块级别划分；另提供流域分离的划分方案（保留4个洲的4个流域用于地理泛化测试）
- 对抗性点放置：每个放置对选自大量候选池（训练/验证96个，测试256个），对模型具有欺骗性，随机放置可导致视线捷径
文件结构与下载
- 主文件：
geopathfinder.tar.gz（下载后解压至geopathfinder/文件夹） - 目录结构：
pathfinder/connect/：标签为1的连接图像（JPEG）pathfinder/disconnect/：标签为0的断开图像（JPEG）images/：无标记点的原始真彩色图像（JPEG）raw/：uint16 RGBN地表反射率GeoTIFF（ZSTD压缩）masks/：水陆掩码（0=陆地，1=水域，2=无效，3=可疑）overlays/：掩码与真彩色融合图像（QC用）- 元数据文件：
metadata.parquet（包含每块的划分、区域、场景ID、边界、CRS/变换、QC统计、点像素及经纬度坐标） - 水多边形文件：
water_polygons.parquet（约281k个Overture水多边形，GeoParquet格式，WGS84） - 其他：
splits.json（划分ID）、snowy_ids.json（雪景排除ID）、plan.json（构建配置快照）
数据来源与许可
- 卫星图像：Landsat Collection 2 Level-2（美国地质调查局，公共领域）
- 水域几何：Overture Maps
base/water数据（开放数据库许可ODbL） - 数据集许可：CC-BY-4.0（注释与图像合成）
快速使用示例
python import pandas as pd meta = pd.read_parquet("metadata.parquet") train = meta[meta.split == "train"]
标签：pathfinder/connect/{sample_id}.jpg (label 1)
pathfinder/disconnect/{sample_id}.jpg (label 0)
基准性能
| 方法 | 测试集准确率 | 放置泛化分数（Easy/Medium/Hard宏平均） |
|---|---|---|
| 无学习的NDWI（绿/近红外） | 69.9% | 76.2% |
| OverLoCK-B（微调） | 88.5±0.5% | 62.4% |
| MaxViT-B | 87.5±0.9% | 63.0% |
| ResNet-50 | 85.3±0.6% | 60.2% |
| ViT-B/16 | 79.3±1.0% | 56.8% |
| ResNet-18（从头训练） | 77.4±0.9% | 53.6% |
- 所有模型在放置泛化分数上表现不佳（模型学会“明显答案往往是错的”），微调后Easy/Medium/Hard三档准确率分布不均，Hard档约75%。
- 若每轮训练重新采样点放置，Macro分数可恢复至85.1±0.3（MaxViT-B），但基准仍未被完全解决。
验证与引用
- 通过
water_polygons.parquet在对应CRS上栅格化可精确复现掩码（IoU ≥ 0.999）。 - 完整基准代码与训练方案：https://github.com/isaaccorley/geo-long-range-arena
- 引用格式： bibtex @misc{corley2026geopathfinder, title = {GeoPathfinder: A Long-Range Spatial Reasoning Benchmark on Real Satellite Imagery}, author = {Corley, Isaac}, year = {2026}, url = {https://huggingface.co/datasets/isaaccorley/GeoPathfinder} }
