Organizations: Advanced Interdisciplinary Institute of Satellite Applications, Beijing Normal University, Beijing 100875, China · State Key Laboratory of Remote Sensing and Digital Earth, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China
Satellite Earth observation has accumulated massive spatiotemporal archives essential for monitoring environmental change, yet these remain organized as discrete raster files, making them costly to store, transmit, and query. We present GeoNDC, a queryable neural data cube that encodes planetary-scale Earth observation data as a continuous spatiotemporal implicit neural field, enabling on-demand queries and continuous-time reconstruction without full decompression. Experiments on a 20-year global MODIS MCD43A4 reflectance record (8016×4008 pixels, 7 bands, 915 temporal frames) show that the learned representation supports direct spatiotemporal queries on consumer hardware. On Sentinel-2 imagery (10 m), continuous temporal parameterization recovers cloud-free dynamics with high fidelity (R2>0.85) under simulated 2-km cloud occlusion. On HiGLASS biophysical products (LAI and FPAR), GeoNDC attains near-perfect accuracy (R2>0.98). The representation compresses the 20-year MODIS archive to 0.44,GB -- approximately 95:1 relative to an optimized Int16 baseline -- with high spectral fidelity (mean R2>0.98, mean RMSE =0.021). These results suggest GeoNDC offers a unified AI-native representation for planetary-scale Earth observation, complementing raw archives with a compact, analysis-ready data layer integrating query, reconstruction, and compression in a single framework.
Figures & tables
Fig. 1: Overview of the GeoNDC representation paradigm . Conventional Earth observation archives are typically organized as discrete, file-based raster tiles or chunked arrays across space and time. GeoNDC reformulates such archives as a queryable spatiotemporal neural data cube, enabling unified point and region queries, arbitrary-time reconstruction, and web-based visualization from a single executable representation.
Method
Native geospatial semantics
Unified spatiotemporal representation
Continuous query support
Selective access
Analysis-oriented design
GeoTIFF / COG
Yes
Partial
No
Yes
Yes
Zarr / NetCDF data cubes
Yes
Yes
No
Yes
Yes
Conventional compression
No
Partial
No
Limited
No
Learned image compression
No
No
No
Limited
No
General INR / neural fields
No
Partial
Yes
Yes
Partial
GeoNDC
Yes
Yes
Yes
Yes
Yes
Table 1: Conceptual positioning of GeoNDC relative to existing geospatial storage, compression, and neural representation paradigms.
Fig. 2: Architecture of the spatiotemporal embedding. A query point (x,y,t) from the original spatiotemporal data cube is processed through a decoupled dual-branch structure designed to address the intrinsic spatiotemporal anisotropy of Earth observations. To preserve high-frequency spatial boundaries, the spatial coordinates (x,y) are queried directly in a static high-resolution 2D HashGrid (bottom). To capture smoother regional spatiotemporal dynamics while reducing temporal striping artifacts, a spatial scaling factor s is applied to generate scaled coordinates (sx,sy,t) , which are then fed into a coarse 3D HashGrid (top). The hierarchical descriptors extracted by interpolation, fxy and fT , are concatenated to form the final spatiotemporal mixed embedding vector.
Fig. 3: The GeoNDC architecture. The framework maps coordinates from the original spatiotemporal data cube into a high-dimensional feature space through spatiotemporal geometric embedding. The resulting mixed embedding vector F is decoded by an MLP to predict the surface state v . Model parameters are optimized through backpropagation by comparing the prediction v with the ground-truth value v . To preserve localized high-frequency details, residuals r=v−v whose magnitude exceeds a threshold τ are quantized and stored in a sparse residual package using entropy coding, forming an optional correction layer on top of the base neural representation.
Fig. 4: The GeoNDC unified storage protocol and access pipeline. The GeoNDC Encoder compresses raw spatiotemporal data cubes into a compact .gndc file composed of three components: (1) a Global Geospatial Header that preserves geospatial metadata required for interoperability, including the coordinate reference system (CRS) and temporal indexing information; (2) a Neural Payload containing the quantized parameters of the learned neural field; and (3) an optional Physical Correction Layer that stores validity masks and sparse residuals. The GeoNDC Reader supports direct random access, allowing arbitrary points or regions in space and time, (x,y,t) , to be queried without full-volume decompression.
Fig. 5: High-resolution surface reconstruction under incomplete observations. (a) Mask-and-Restore experiment: Visual comparison between Sentinel-2 snapshots with simulated gaps of varying scales (small, medium, and large; left) and the corresponding GeoNDC reconstructions (right). (b) Reconstruction fidelity in valid regions: Density scatter plots for the Red (B4) and NIR (B8) bands, comparing GeoNDC reconstructions with the original observations in non-masked regions. (c) Recovery accuracy across gap sizes: Quantitative comparison between GeoNDC reconstructions and the ground-truth observations within masked regions for different gap scales. (d) Interpolation baseline comparison: Reconstruction performance of a traditional linear interpolation method over the same masked regions and spectral bands.
Method
Size (MB)
Ratio
R2
RMSE
GeoNDC (amortized)
0.52
3 486:1
0.9943
0.0199
COIN-L (265k params)
0.53
3 385:1
0.9535
0.0543
COIN-M (67k params)
0.13
13 360:1
0.9445
0.0589
COIN-S (9k params)
0.02
100 322:1
0.9047
0.0769
Int16 + Zstandard
80.68
22:1
1.0000
0.0000
JPEG2000 (Q=20)
61.83
30:1
1.0000
0.0007
Table 2: Single-frame compression comparison on global MODIS snapshots (mean over four dates spanning 2021). Per-frame size for GeoNDC is the total model size amortized over all 915 temporal frames.
Fig. 6: Temporal gap-filling in a cloud-affected sub-region (southern China). Top: Two-year NDVI time series at a representative vegetation pixel. Pink shading indicates cloud-masked observation dates. GeoNDC (red curve) closely tracks the original observations (dark line) where data are available and provides continuous reconstruction through cloud-masked gaps. The seasonal phenological cycle is preserved without external gap-filling. Bottom: Spatial comparison on January 30, 2022, when only 16% of pixels are valid. The original NDVI map (left) shows extensive data loss due to winter cloud cover, while GeoNDC (right) reconstructs a spatially complete vegetation field.
Fig. 7: Planetary-scale compression and queryability. (a) Storage reduction at planetary scale: Storage comparison for the 20-year global MODIS archive. GeoNDC (0.44 GB) achieves an approximately 380:1 compression ratio relative to the raw double-precision archive (168 GB), and an approximately 95:1 ratio relative to a theoretical Int16 raster baseline (42 GB). (b) Interactive query and visualization: The compressed model supports direct spatiotemporal access on consumer hardware (RTX 4080). The interface shows interactive retrieval of a multi-year phenological trajectory (inset) at a selected pixel in the Amazon. (c) Surface state reconstruction: A global snapshot of the reconstructed NDVI field on June 24, 2007, showing the large-scale vegetation state during Northern Hemisphere summer. (d) Illustrative temporal derivative field: An instantaneous temporal derivative field ( ∂NDVI/∂t ), computed from the learned neural representation via automatic differentiation. The map shows coherent latitudinal patterns consistent with expected phenological gradients, illustrating a capability inherent to continuous neural representations.
Fig. 8: Shared neural representation of multi-product biophysical variables (LAI and FPAR). (a) Joint representation workflow: Two large discrete GeoTIFF archives (2.32 GB for LAI and 4.88 GB for FPAR) are jointly parameterized into a single compact GeoNDC model (385 MB) through a shared spatiotemporal representation. (b) Interactive product serving: The GeoNDC Viewer interface demonstrates direct rendering and exploration of the compressed high-resolution ( 20m ) HiGLASS tile. (c) Quantitative reconstruction accuracy: Density scatter plots for 4,000 randomly sampled points across diverse seasons show near-perfect agreement between the original and reconstructed values for both LAI ( R2=0.9967 ) and FPAR ( R2=0.9888 ). (d) Temporal consistency: Comparison of multi-year temporal trajectories shows that the reconstructed series preserves the seasonal amplitude, peak vegetation states, and fine temporal oscillations without introducing obvious smoothing artifacts.
Query type
GeoTIFF
GeoNDC
Speedup
Single-pixel time series (1 px × 46 steps) a
612 ms
8 ms
81 ×
Regional subset (300 2 × 46 steps)
2.9 s
472 ms
6.2 ×
Per-frame memory (loaded)
1.7 GB RAM (1 day)
2.2 GB VRAM b (20 yr)
On-disk size
8.7 GB (1 yr)
264 MB (20 yr)
a GeoNDC returns all 915 time steps (20 yr) in 8 ms; GeoTIFF retrieves only the 46 steps of 2021.
b Full 20-year model loaded once; provides access to all 915 frames.
Table 3: Query efficiency comparison between GeoNDC and conventional GeoTIFF raster access on the 2021 MODIS MCD43A4 archive (46 temporal frames). GeoNDC serves all queries from a single neural model covering the full 20-year archive (915 frames), preloaded into GPU memory. GeoTIFF queries read from individual float64 raster files stored on a local NVMe SSD.
In this work, we present LIANet (Location Is All You Need Network), a coordinate-based neural representation that models multi-temporal spaceborne Earth observation (EO) data for a given region of interest as a continuous spatiotemporal neural field. Given only spatial and temporal coordinates, LIANet reconstructs the corresponding satellite imagery. Once pretrained, this neural representation can be adapted to various EO downstream tasks, such as semantic segmentation or pixel-wise regression, importantly, without requiring access to the original satellite data. LIANet intends to serve as a user-friendly alternative to Geospatial Foundation Models (GFMs) by eliminating the overhead of data access and preprocessing for end-users and enabling fine-tuning solely based on labels. We demonstrate the pretraining of LIANet across target areas of varying sizes and show that fine-tuning it for downstream tasks achieves competitive performance compared to training from scratch or using established GFMs. The source code and datasets are publicly available at https://github.com/mojganmadadi/LIANet/tree/v1.0.1.
Mojgan Madadikhaljan, Jonathan Prexl, Isabelle Wittmann +2
University of the Bundeswehr Munich · IBM Research – Europe · Columbia University +1
Satellite observations, precomputed embeddings, and map products describe the same evolving Earth, yet are stored as independent, petabyte-scale data products. Their continued growth calls for compact representations of multiple products while preserving spatial and temporal detail. We introduce Planetary Feature Fields (PFFs), which exploit redundancy across data products by modeling them jointly as continuous functions of space and time at planetary scale. PFFs are spatially local explicit-implicit (hybrid) neural fields. Each field shares a factored feature volume---a decomposition of an explicit 3D grid with smaller factors---across products, while lightweight implicit decoders reconstruct individual products across multiple timesteps. PFFs reconstruct EO products over space and time more accurately than single-product fields at matched compression rates. At 1800× compression relative to the uncompressed source data, reconstructed features retain approximately 90% or more of the performance achieved with the original features on pixel-level segmentation, change detection, and patch-level classification tasks. PFFs can add new timesteps by extending their factored feature volumes and add new products by attaching new decoders, while leaving existing outputs unchanged. PFFs reduce end-to-end feature access latency by an order of magnitude relative to evaluated API and cloud-storage pipelines.
Arjun Rao, Sebastian Loeschcke, Anthony Fuller +3
University of British Columbia · University of Copenhagen · Carleton University +2
Earth observation is becoming one of the largest data-producing activities in science, yet current pipelines still treat compression as a storage and transmission tool rather than a new way to use data. We present a generative compression framework that learns from historical Earth observation archives and enables on-demand 100x to 10,000x data reduction across downstream tasks. Unlike general visual data, Earth observation repeatedly measures the same evolving planet, making historical-prior learning feasible for extreme compression. To realize this paradigm, we train large generative compression models at exascale on the LineShine Armv9 CPU supercomputer, with co-optimization across model design, kernels, memory hierarchy, runtime, and parallelism. Our implementation sustains 1.54 EFLOP/s and peaks at 2.16 EFLOP/s in end-to-end training. This work shows that historical-prior generative compression can turn Earth observation data into an active, task-adaptive foundation for acquisition, delivery, storage, and scientific use.
Jinxiao Zhang, Runmin Dong, Xiyong Wu +14
Department of Earth System Science, Tsinghua University · Sun Yat-Sen University · Institute of Data and Information, Tsinghua Shenzhen International Graduate School +2