GeoNDC: A Queryable Neural Data Cube for Planetary-Scale Earth Observation

Jianbo Qi, Mengyao Li, Baogui Jiang, Yidan Chen, Xihan Mu, Qiao Wang, [4pt] Advanced Interdisciplinary Institute of Satellite Applications, Beijing Normal University Beijing 100875, China · 2026

Read GeoNDC: A Queryable Neural Data Cube for Planetary-Scale Earth Observation by Jianbo Qi, Mengyao Li, Baogui Jiang, Yidan Chen, Xihan Mu, Qiao Wang, [4pt] Advanced Interdisciplinary Institute of Satellite Applications, Beijing Normal University Beijing 100875, China (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-25037.txt.

Introduction

Qi and co-authors pack twenty years of Earth scans into a queryable neural cube called GeoNDC. Satellite archives sit as stacks of raster files, which are heavy to store, move, and search. GeoNDC instead learns a smooth field over space and time that can answer point queries on plain hardware. On the global MODIS record it shrinks the archive about 95 to 1, down to 0.44 gigabytes, with high fidelity. On Sentinel-2 views it rebuilds clear scenes through fake cloud gaps. Plant trait products come back with near-perfect scores.

Catalog record

CREATORJianbo Qi, Mengyao Li, Baogui Jiang, Yidan Chen, Xihan Mu, Qiao Wang, [4pt] Advanced Interdisciplinary Institute of Satellite Applications, Beijing Normal University Beijing 100875, China
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63123
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.25037

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GeoNDC: A Queryable Neural Data Cube for Planetary-Scale Earth Observation Jianbo Qi, Mengyao Li, Baogui Jiang, Yidan Chen, Xihan Mu, Qiao Wang, [4pt] Advanced Interdisciplinary Institute of Satellite Applications, Beijing Normal University Beijing 100875, China Abstract 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 8016\times 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 ( R 2 > 0.85 R^{2}>0.85 ) under simulated 2-km cloud occlusion. On HiGLASS biophysical products (LAI and FPAR), GeoNDC attains near-perfect accuracy ( R 2 > 0.98 R^{2}>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 R 2 > 0.98 R^{2}>0.98 , mean RMSE = 0.021 =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. Keywords: Neural Data Cube, Implicit Neural Representation, Earth Observation, Queryable Representation, Spatiotemporal Modeling, Geospatial Data Cube Subjects: Physics - Geophysics (physics.geo-ph);Computer Science - Computer Vision and Pattern Recognition (cs.CV) 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. 1 Introduction Over the past decades, Earth observation (EO) has evolved into a foundational infrastructure for monitoring planetary change [ 30 ] . Long-term satellite missions such as MODIS and Landsat have accumulated massive spatiotemporal archives, enabling the quantification of processes ranging from global greening to ecosystem resilience loss [ 11 , 29 , 3 ] . Yet the central challenge of these archives is no longer data acquisition alone. At planetary scale, EO systems increasingly face a coupled bottleneck of storage, access, query, and analysis: vast data volumes must not only be preserved, but also transmitted efficiently, queried interactively across space and time, and transformed into scientifically meaningful indicators [ 5 , 17 ] . As sensor constellations continue to improve revisit frequency, spatial resolution, and spectral richness, this gap between data availability and data usability is becoming a major constraint on Earth system science. Although substantial progress has been made in cloud-native geospatial infrastructures, learned compression, and neural representations, these advances still address the problem only in partial form. Existing archive systems make EO data easier to store and distribute, while compression methods reduce file size and neural fields provide continuous signal parameterizations. However, none of these directions directly yields a planetary-scale EO representation that is simultaneously compact, geospatially queryable, and continuous in time. As a result, storage, retrieval, and reconstruction remain largely separated stages in current EO workflows. This limitation becomes especially pronounced for long-term geospatial data cubes. Many land-surface processes are continuous in both space and time, yet satellite observations remain discrete, incomplete, and frequently interrupted by clouds, shadows, and irregular sampling. In practice, this forces downstream analysis to rely on compositing, masking, or interpolation [ 14 ] . What is missing is not simply a better codec or a faster archive format, but a representation in which compact storage, direct query, and continuity-aware reconstruction are unified by design. Motivated by this gap, we present GeoNDC (Geographic Neural Data Cube), a queryable neural data cube for planetary-scale Earth observation. GeoNDC represents georeferenced EO observations as a continuous spatiotemporal neural field conditioned on spatial coordinates and time, allowing the archive to be stored not merely as raster files, but as a compact executable model. In this formulation, querying a pixel value becomes function evaluation and temporal interpolation becomes native continuous-time inference. GeoNDC is therefore not merely a compression tool; it is an AI-native representation and access paradigm for EO data cubes that unifies storage, query, and reconstruction within a single framework. Building on multi-resolution hash encodings and a decoupled spatiotemporal architecture, GeoNDC is designed to preserve the practical semantics of EO archives while changing their underlying representation. As a lossy neural representation, GeoNDC is not intended to replace authoritative raw EO archives; rather, it provides an analysis-ready and AI-ready layer for compact access and downstream scientific use. It supports geospatially aware random access, continuous reconstruction under missing observations, and compact deployment on standard hardware. This shift from data as files to data as models provides a new perspective on planetary EO archives: instead of distributing ever-growing collections of discrete rasters, we can learn portable neural data cubes that retain the dominant spatiotemporal structure of the observations while remaining directly queryable. The main contributions of this study are as follows. First, we propose GeoNDC, a queryable neural data cube that reformulates planetary-scale EO archives as continuous spatiotemporal neural representations. Second, we show that this representation achieves orders-of-magnitude compression while preserving the dominant spatial and temporal dynamics of long-term satellite observations. Third, we demonstrate that the learned neural cube supports temporally coherent reconstruction under missing observations. Together, these results suggest a new route toward analysis-ready and AI-ready EO infrastructure, in which compression, access, and reconstruction are integrated into a unified representational framework. 2 Related Work 2.1 Earth observation data cubes and cloud-native geospatial infrastructures A long-standing challenge in Earth observation is that the scientific value of satellite archives is often constrained not only by data volume, but also by the difficulty of organizing, accessing, and repeatedly analyzing large spatiotemporal collections. Earth observation data cubes (EODCs) were developed precisely to address this problem by transforming large image archives into analysis-ready, spatiotemporally aligned data structures that facilitate repeated querying and large-scale computation [ 15 , 16 , 9 , 1 ] . Related work has further emphasized interoperability, view-based architectures, and cloud-native access patterns as essential ingredients for scalable Earth data systems [ 22 , 10 , 2 ] . These developments have significantly improved the practicality of large EO archives. However, they remain predominantly file-centric : even when organized as data cubes or cloud-optimized rasters, the underlying representation is still a collection of discrete arrays to be stored, streamed, opened, and decoded before analysis. In other words, current EODC and cloud-native infrastructures substantially improve how EO data are managed and accessed, but do not fundamentally change what the archive is represented as. GeoNDC is motivated by this distinction. Rather than reorganizing the archive into a more efficient raster-centric structure, it seeks to represent the archive itself as a compact neural data cube that is directly queryable in space and time. 2.2 Implicit neural representations and learned compression In parallel with advances in geospatial data infrastructure, computer vision and machine learning have developed a rich literature on implicit neural representations (INRs),

Read the complete text