Geographic Neural Data Cube
Compress decades of satellite data into a tiny executable model. Query any point on Earth, instantly.
20-year global MODIS · 168 GB → 0.44 GB · 380× compression
数据即模型，读取即推理 — Data as Model, Read as Inference
Encode massive remote sensing archives into tiny neural models. Achieve 10× to 100×+ compression—terabytes of satellite imagery distilled into sub-gigabyte executables that run anywhere.
Represent spatiotemporal remote sensing data as geo-referenced neural data cubes. Query any coordinate on demand—no full reconstruction required. A native AI representation of the Earth.
Spatiotemporal data lives on an implicit neural field—a continuous differentiable manifold. Compute derivatives like dNDVI/dt via automatic differentiation. Enable ecological velocity analysis at any scale.
Compression performance on representative datasets
7 spectral bands, 5 km resolution, 20 years of global coverage. The core dataset used in our paper experiments.
Leaf Area Index and FPAR for mainland China. 20 m resolution, 5-day temporal resolution.
Compressed earth observation data ready for download
Leaf Area Index and FPAR for mainland China. 20m resolution, 5-day temporal resolution, 2018-2023.
Global MODIS data cube covering 2003-2023. 7 spectral bands, 5km resolution, the core dataset for paper experiments.
High-resolution Sentinel-2 data for China region. 10m resolution, 500+ scenes planned for 2024-2025.
Peer-reviewed research on neural earth observation
Jianbo Qi, Mengyao Li, Baogui Jiang, Yidan Chen, Qiao Wang
We present GeoNDC, a neural earth data cube that compresses 168 GB of global MODIS imagery (2003-2023) into just 0.44 GB while preserving R2 > 0.85 reconstruction accuracy. By learning implicit neural representations of planetary-scale vegetation dynamics, GeoNDC enables instant queries of any location on Earth without decompression. Our approach introduces phenological inpainting for cloud removal and computes ecological derivatives through automatic differentiation, achieving what we call "planetary calculus."
Project website and initial datasets released
70 scenes completed, 1139 total planned
Python SDK for model loading and querying