XuanNv Tech Launches 'XuanNv Earth' Geospatial Foundation Model
On September 24, Beijing XuanNv Technology Co., Ltd. officially released its self-developed Earth-embedded spatiotemporal intelligence foundation model, "XuanNv Earth," along with a China-region embedding vector dataset. The model is built on long-sequence, multi-modal remote sensing data and achieves unified surface embedding representation of geoscience knowledge at a 10-meter grid, providing a domestic AI-powered data foundation for natural resource supervision, ecological protection and restoration, disaster prevention and mitigation, and carbon sink accounting.
The release includes two core deliverables: the foundation model and the embedding vector dataset. The foundation model acts as a production tool to extract general surface features from remote sensing imagery, while the embedding vector dataset is a ready-to-use data product generated by the model for downstream industry applications.
Compared with Google's AEF Earth foundation model, XuanNv Earth follows a similar technical route of encoding multi-source, multi-temporal remote sensing data into surface embedding vectors, but differs in positioning. It focuses on China, providing quarterly embedding vectors and planning monthly products. It also integrates more domestic satellite resources, including meter-level high-resolution data, enabling finer depiction of ground objects and geoscience processes.
According to the company, the model mines spectral, texture, morphological, spatial and temporal patterns from massive remote sensing images and converts them into reusable embedding vectors. Downstream applications in agriculture, natural resources, urban governance and ecology can train task-specific heads with minimal labeled samples to quickly generate thematic layers such as cropland, water bodies, buildings and roads, reducing development barriers and business costs.
Industry experts say the geospatial embedding approach helps address issues such as data overload, inconsistent standards and cross-region model transfer difficulties in ecological monitoring, and can serve as a spatial-temporal basis for natural resource asset accounting.
