AiTLAS: Multi-Task Satellite Imagery Benchmarks for Earth Observation

AiTLAS is a collection of satellite imagery datasets and benchmarks for Earth observation, developed by biasvariancelabs and updated in March 2026. It provides labeled Sentinel satellite data for training and evaluating deep learning models across classification, detection, and segmentation tasks.

Use Cases

• Training object detection models to identify structures in Sentinel satellite imagery
• Executing semantic segmentation for land cover classification using geospatial masks
• Benchmarking Vision Transformer performance on remote sensing image classification

Strengths

• Supports multi-task learning including classification, detection, and segmentation
• Apache-2.0 open-source license
• Optimized for PyTorch and Vision Transformer architectures

Limitations

• Geographic coverage and temporal range of the imagery are not explicitly defined
• Lack of detailed metadata regarding the specific number of classes or labels

Provenance

Source

biasvariancelabs

Collection Method

Satellite sensor data acquisition

Freshness

Updated as of March 2026.