Awesome Satellite Imagery Datasets

Dataset Overview

Instance Segmentation

• Spacenet Challenge Round 4 - Off‑nadir
  The dataset contains 126 k building footprints (Atlanta), 27 WorldView‑2 images (0.3 m resolution) captured at off‑nadir angles ranging from 7° to 54°. Images are resampled using bicubic interpolation to a uniform pixel count to compensate for coarse native resolution at high off‑nadir.

• Airbus Ship Detection Challenge
  Contains 131 k ships, 104 k training / 88 k testing image tiles, satellite imagery (1.5 m resolution). Raster mask labels are encoded in run‑length format; Kaggle kernels are available.

• Open AI Challenge: Tanzania
  Provides building footprints and three building condition categories, RGB drone imagery; data links are available in a Google Sheet.

• Netherlands LPIS agricultural field boundaries
  294 crop/vegetation classes, 780 k parcels, yearly datasets from 2009‑2018.

• Denmark LPIS agricultural field boundaries
  293 crop/vegetation classes, 600 k parcels, yearly datasets from 2008‑2018.

• CrowdAI Mapping Challenge
  Provides building footprints, RGB satellite imagery, COCO format.

• Spacenet Challenge Round 2 - Buildings
  685 k building footprints, 3/8‑band WorldView‑3 images (0.3 m resolution), 5 cities, SpaceNet Challenge asset library.

• Spacenet Challenge Round 1 - Buildings
  Building footprints (Rio de Janeiro), 3/8‑band WorldView‑3 images (0.5 m resolution), SpaceNet Challenge asset library.

Object Detection

• DOTA: Large‑scale Dataset for Object Detection in Aerial Images
  15 object categories, 188 k instances, Google Earth image chips, Faster‑RCNN benchmark (MXNet), academic‑only use.

• xView 2018 Detection Challenge
  60 object categories, 1 million instances, WorldView‑3 images (0.3 m resolution), COCO format, pretrained TensorFlow and PyTorch benchmarks.

• Open AI Challenge: Aerial Imagery of South Pacific Islands
  Tree locations and four tree species, RGB drone imagery (0.4 m / 0.8 m resolution), multiple AOIs in Tonga.

• NIST DSE Plant Identification with NEON Remote Sensing Data
  Tree locations, species, canopy parameters; hyperspectral (1 m) and RGB (0.25 m) imagery, LiDAR point clouds, canopy height models.

• NOAA Fisheries Steller Sea Lion Population Count
  5 sea‑lion classes, ~80 k instances, ~1 k aerial images, Kaggle kernels.

• Spacenet Rio De Janeiro Points of Interest Dataset
  460 object categories, 120 k points (11 k manually verified), 3/8‑band WorldView‑3 images (0.5 m resolution), SpaceNet Challenge asset library.

• Cars Overhead With Context (COWC)
  32 k vehicle bounding boxes, aerial imagery (0.15 m resolution), 6 cities.

Semantic Segmentation

• Agricultural Crop Cover Classification Challenge
  Two primary classes: corn and soybean, Landsat‑8 imagery (30 m), USDA crop data layer as ground truth.

• Spacenet Challenge Round 3 - Roads
  8 000 km of roads, 5 city AOIs, 3/8‑band WorldView‑3 images (0.3 m resolution), SpaceNet Challenge asset library.

• Urban 3D Challenge
  157 k building footprints, RGB orthophotos (0.5 m resolution), DSM/DTM, 3 cities, SpaceNet Challenge asset library.

• DSTL Satellite Imagery Feature Detection Challenge
  10 land‑cover classes, 57 1 × 1 km images, 3/16‑band WorldView‑3 imagery (0.3 m‑7.5 m resolution), Kaggle kernels.

• Inria Aerial Image Labeling
  Semantic segmentation (buildings), RGB aerial imagery (0.3 m), 5 cities.

• ISPRS Potsdam 2D Semantic Labeling Contest
  Six urban land‑cover classes, raster mask labels, 4‑band RGB‑IR aerial imagery (0.05 m) and DSM, 38 image chips.

Chip Classification (Image Recognition)

• Statoil/C‑CORE Iceberg Classifier Challenge
  Two classes: ships and icebergs, 2‑band HH/HV polarimetric SAR imagery, Kaggle kernels.

• Functional Map of the World Challenge
  63 classes, 1 million chips, 4/8‑band satellite imagery (0.3 m), COCO format, benchmark models.

• Planet: Understanding the Amazon from Space
  13 land‑cover classes + 4 cloud‑condition classes, 4‑band (RGB‑NIR) satellite imagery (5 m), Amazon rainforest, Kaggle kernels.

• Deepsat: SAT‑4/SAT‑6 airborne datasets
  6 land‑cover classes, 400 k 28 × 28‑pixel chips, 4‑band RGB‑NIR aerial imagery (1 m), stored as .mat files.

• UC Merced Land Use Dataset
  21 land‑cover classes, 100 chips per class, aerial imagery (0.30 m).

Other / Multi‑Task

• DEEPGLOBE – 2018 Satellite Challenge
  Three tracks: road extraction, building detection, land‑cover classification.

• IEEE Data Fusion Contest 2018
  20 land‑cover classes, fused data sources: multispectral LiDAR, hyperspectral (1 m), RGB imagery (0.05 m).

• TiSeLaC: Time Series Land Cover Classification Challenge
  Land‑cover time‑series classification (9 classes), Landsat‑8 (23 temporal images, 10 bands, 30 m), Réunion Island.

• Multi‑View Stereo 3D Mapping Challenge
  Develop MVS 3D‑mapping algorithms converting high‑resolution WorldView‑3 satellite images to 3D point clouds; 0.2 m LiDAR ground truth.

• Draper Satellite Image Chronology
  Predict the chronological order of images taken at the same location within a 5‑day window; Kaggle kernels are provided.