Remote sensing data has long been plagued by the pain points of "high procurement cost, heavy preprocessing workload, and great difficulty in reuse": Xuannu Earth applies 10-meter grid surface embedding vectors, which turns the traditional logic of "preparing dedicated data for every application" into "one-time embedding for multi-task reuse".
On September 24, 2026, Xuannv Technology launched Xuannv Earth, a self-developed geospatial intelligent base large model with embedded Earth capabilities, and simultaneously released the supporting embedded vector dataset covering China's territory. With a 10m grid and quarterly update frequency, the model encodes multi-source remote sensing observation data into reusable surface embedding vectors, aiming to solve the long-standing pain points in remote sensing applications of "high data purchase cost, heavy preprocessing workload and low reusability".
Remote sensing data features "high acquisition volume but low utilization rate", downstream applications are restricted by high costs
The downstream applications of remote sensing and geographic information have long been constrained by the cost and efficiency of data links. Cao Bingxuan, Director of the Digital Intelligence Ecological Innovation Center of China Urban Construction Research Institute Co., Ltd., noted in public comments that the common industry model in the past was that customers purchased various isolated data on demand at high prices in a decentralized manner, leading to high procurement costs. A large amount of additional manpower had to be invested in subsequent data cleaning and fusion processing, resulting in a long overall decision-making chain and high comprehensive costs.
Behind this cost structure lies a chain of "application determines model, model determines data, data determines cost": once the application requirements change, both the model and data need to be adjusted accordingly, keeping the costs of data procurement, model R&D and business adaptation at a high level with low reusability. For scenarios such as ecology and natural resources, there are also problems of inconsistent data standards and difficult cross-regional model migration, and the same set of methods often needs to be restarted from scratch when applied from one region to another. Yu Min, Head of the Jiangxi Regional Center of China Urban Construction Research Institute Co., Ltd. and professor-level senior engineer, pointed out that data overload, inconsistent standards and difficult cross-regional model migration in the ecological field are practical problems that have long restricted industry efficiency.
The demand side is also expanding simultaneously. The *China Geographic Information Industry Development Report (2025)* released by China Association for Geographic Information Industry shows that the total output value of China's geographic information industry reached 850.1 billion yuan in 2024, with a growth rate of 4.8%, and there are more than 24,000 operating entities; the association estimates that this output value will increase to nearly one trillion yuan in 2025. At the policy level, the surveying and mapping geographic information cause under the 15th Five-Year Plan is accelerating its digital and intelligent transformation, and the Ministry of Natural Resources has proposed to promote the in-depth integration of artificial intelligence into the whole chain of data production and update. The intelligent version (2026) of Tianditu, the national geographic information public service platform, has also been officially launched on September 24, 2026. Businesses including cultivated land protection, ecological restoration, disaster prevention and mitigation, and carbon sink accounting have put forward higher requirements for high-frequency, standardized and reusable surface data.
One-time embedding, multi-task reuse, making surface features a callable base
The solution provided by Xuannv Earth is to advance the process of "preparing data and training models separately for each application". Its core technology is independently developed by a domestic team, adopting the technical system of "Earth embedding + multi-source data alignment". Based on the long-time series and multi-modal remote sensing observation data including China's high-resolution satellites and commercial satellites, it realizes the unified geoscience knowledge surface embedding representation that integrates meter-level data features under the 10m grid.
In terms of products, Xuannv Technology has released two achievements this time. The first is the Xuannv Earth base large model, which as the core "production tool", is responsible for extracting general surface features from remote sensing images; the second is the supporting embedded vector dataset, as the directly implementable "data product", generated by processing the original images with the model. The first-phase product covers quarterly embedded vector data from 2020 to 2021, which is managed and served by 10 partitions across China, and monthly-level products will be launched successively. The model converts the spectral, textural, morphological, spatial correlation and time-series evolution rules contained in massive remote sensing images into a standardized, infinitely reusable embedded vector set at one time.
For downstream users, the real change lies in the development threshold. In the past, users needed to complete preprocessing steps such as radiometric correction, geometric registration, cloud and stripe removal by themselves, and then label a large number of samples to train the model; now, users only need to train the "decoding head" for downstream tasks with a very small number of business-labeled samples on the base, to quickly output thematic result layers such as cultivated land, water bodies, buildings and roads.
Compared with similar international products, Xuannv Earth is more focused on the Chinese market. Google's AlphaEarth Foundations (AEF) focuses on the global scope and outputs annual-level embedded vectors; Xuannv Earth provides quarterly embedded vectors and integrates more domestic satellite resources, including a large number of higher-precision meter-level satellite data, which can describe ground objects and geoscience processes in a more refined way. The feasibility of this technical route has been verified in regional practices before: in the public practice of "Xuannv Earth" on ModelScope, the team verified the sparse annotation mapping path of "a small number of annotations + lightweight task head" that can quickly adapt to downstream tasks such as buildings, roads, water bodies and green spaces, based on 10m, 64-dimensional monthly embeddings of 320 sample blocks in Haidian District, Beijing.
From the base to the "decoding head", commercialization path and team progress
In terms of business model, Xuannv Technology is trying to reconstruct the cost relationship of "data - model - application", adopting the mode of "lightweight on-demand procurement + centralized large model processing": customers no longer need to pay for scattered redundant data, nor do they need to undertake subsequent complex data processing work. The target users cover application parties in multiple industries such as agriculture, natural resources, urban governance and ecological environment, especially small and medium-sized teams and local institutions that have industry insights but cannot afford the cost of the entire data system and model R&D.
In the ecological and natural resources scenarios, this base has more specific landing points. Yu Min believes that geographic embedding technology converts multi-source heterogeneous Earth observation data into a unified, compact vector representation, which can alleviate the problems of data overload, inconsistent standards and difficult cross-regional model migration in the ecological field. It can not only support dynamic ecological monitoring, but also serve as the physical quantity spatiotemporal base for natural resource asset accounting, dynamically updating the distribution, area and change patches of assets such as land, forest and grass, wetlands and water bodies, providing basic data support for the compilation of natural resource asset balance sheets, GEP physical quantity accounting and supervision and audit.
Xuannv Technology was founded on March 25, 2026, registered at Dongsheng Building, No. 8 Zhongguancun East Road, Haidian District, Beijing, with a registered capital of 5 million yuan; according to public industrial and commercial registration information, the company's legal representative is Nian Jie, and its core technology is independently developed by a domestic team. At this stage, the company has officially released the Xuannv Earth base large model and the China-region embedded vector dataset on September 24, 2026, and opened the first batch of data simultaneously. The R&D team said that it will cooperate with universities, scientific research institutes and industrial partners in the future to promote the large-scale application of Earth intelligence technology in actual business. In terms of financing, the company is advancing its seed round financing.
From "data buyout" to "vector reuse", Xuannv Earth has chosen a more underlying path. Whether it can stably convert the "China-region, quarterly-level" embedding capability into an industry-scalable deliverable, and enable downstream users to obtain usable results at extremely low annotation costs, will be the issue that this new company needs to continuously address after the seed round.