I previously shared an evaluation of remote sensing foundation models by an overseas expert in “Evaluating 89 Remote Sensing Foundation Models: Most Are Not Ready for Direct Use”. Some readers asked whether I could evaluate models from China. In fact, models in China are mostly closed-source and not publicly available, so such an evaluation is not feasible. However, China has released quite a few spatiotemporal foundation models. What are their characteristics and main application areas? Today I will provide an overview.

Kongtian Lingmou

Developed by the Aerospace Information Research Institute, Chinese Academy of Sciences, and first released in August 2022, Kongtian Lingmou is China's first generative pre-trained model for cross-modal remote sensing data. Version 3.0 was released in September 2024 at the Jinan Aerospace Information Technology Conference, with parameter scale reaching ten billion. It covers visible light, SAR, thermal infrared, multispectral, as well as aerial aircraft and UAV data, and supports scene classification, object detection, semantic segmentation, and change detection. It is currently applied mainly in emergency management, land, ocean, and housing and construction.

Kunyuan

Developed by the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, together with the Institute of Tibetan Plateau Research and the Institute of Automation, and released in Beijing in September 2024, Kunyuan is positioned as the world's first multimodal geographic science foundation model. “Kun” refers to the earth, and “Yuan” means both origin and neuron, so the name means the earth's neural system. On geographic science benchmark test sets, its accuracy is 31.3% higher than that of general-purpose language models. In July 2025, the team released two derivative models at the Yantai Coastal Zone Conference: “Kunyuan·Gan Dongnan” for land resource and environmental monitoring in Southeast Asia, and “Kunyuan·Ni Qianhai” for modeling kilometer-scale layered ocean flow fields.

Mogan Xuanyan

Developed by Mogan Mountain Geospatial Information Laboratory and released in Deqing, Zhejiang in January 2025, it is the industry's first hundred-billion-parameter multimodal geospatial foundation model. The model comprises a geospatial language foundation model, a vision foundation model, and a multimodal foundation model, trained on hundreds of millions of multimodal samples. Version 1.0 supports six major capabilities: remote sensing visual perception, spatiotemporal data query and analysis, domain-specific knowledge Q&A, automatic report generation, intelligent task orchestration, and tool calling. It is applied in scenarios such as natural resource satellite image enforcement and geological disaster risk prevention and control. On July 20, 2025, it was upgraded to version 1.5 with support for 3D generative modeling. In July 2026, the National AI Application Pilot Base was inaugurated in Deqing, and it was announced that Mogan Xuanyan went online and began providing services externally.

Zhongce Kunxi Series

Developed by the Chinese Academy of Surveying and Mapping, it was first released collectively at a scientific and technological achievements promotion meeting in December 2025 and later named the Zhongce Kunxi Series. It is not a single model. Zhongce Kunyu performs AI geodesy to serve the national comprehensive spatiotemporal system. Zhongce Kunyuan focuses on engineering-oriented remote sensing interpretation. Zhongce Kunshi performs interactive understanding of remote sensing imagery. Zhongce Kunjie performs map recognition and inspection.

Linlong

Developed by the Smart Forestry and Grassland Innovation Team of the Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, and announced around April 2025, Linlong is positioned as China's first foundation model for the forestry and grassland industry. Its capabilities include tree species recognition, phenotypic parameter extraction, wildlife recognition, pest and disease monitoring, forest fire identification, ecosystem assessment, and management decision-making, with tree species recognition accuracy exceeding 93%. It has already been put into use in the Three-North Shelterbelt Forest Program demonstration areas, national parks, and state-owned forest farms.

Panyu

Developed by the Sichuan Earthquake Administration together with the Sichuan Provincial Bureau of Surveying, Mapping and Geoinformation and the Ya'an Municipal People's Government, and released in Chengdu in August 2026, Panyu is a vertical industry foundation model for earthquake disaster risk prevention. It covers the full chain of earthquake disaster prevention, reduction, and relief, focusing on pre-earthquake risk prevention and control and post-earthquake golden 72-hour emergency response. In the regular phase, it relies on a 3D one-map system for risk sensing and hidden hazard monitoring. During earthquakes, it quickly assesses the disaster situation, generates thematic maps, and formulates response plans. It has already been applied in Shimian County, Ya'an.

Xingtong

Developed by GEOVIS, Xingtong serves as the core of the GEOVIS Cloud Intelligent Computing platform and is positioned as a ten-billion-parameter space-air-ground integrated multimodal foundation model. It uses a mixture-of-experts strategy to integrate optical, SAR, hyperspectral, and UAV imagery, and then performs cross-modal association with natural language and geographic text. Its external product is the GEOVIS Earth Intelligence Engine, which provides remote sensing processing, intelligent interpretation, 3D reconstruction, natural language query, and report generation.

SuperMap LIM

Developed by SuperMap and launched as a commercial product in 2024, SuperMap LIM is integrated into the SuperMap AIF geospatial AI foundation for image feature extraction and change detection. There is also an AI assistant on the desktop side and AgentX on the server side, making it easier for existing SuperMap users to integrate it into current GIS workflows.

SkySense

Developed by the joint laboratory of the School of Remote Sensing and Information Engineering, Wuhan University, and Ant Group, SkySense was released at the World Artificial Intelligence Conference in July 2024. It has 2 billion parameters and covers 17 datasets for scene classification, semantic segmentation, change detection, and crop classification. It is applied in forest conservation, green finance, and agricultural monitoring.

DAMO Academy Remote Sensing Foundation Model

Developed by Alibaba DAMO Academy and released in October 2023 under the model name AIE-SEG, it is positioned as a general-purpose segmentation model for remote sensing AI interpretation, supporting satellite and UAV imagery. Its instance extraction accuracy is about 25% higher than traditional remote sensing models, and change detection is about 30% higher. The Shandong Provincial Institute of Land Surveying and Mapping used it to monitor winter wheat growth across the province, achieving recognition accuracy above 90%. The entry point is the AI Earth Geoscience Cloud Platform, which provides a toolbox and APIs.

SenseEarth (SenseTime Dijie)

Developed by SenseTime, SenseEarth 3.0 was released at WGDC in 2023. I previously introduced it in “SenseTime Releases SenseEarth 3.0 Intelligent Remote Sensing Cloud Platform”. Its public capabilities cover parcel segmentation, object detection, change detection, and super-resolution. Parcel segmentation achieves an average accuracy above 80% on a million-scale parcel validation set, and change detection has been used in natural resource law enforcement and supervision in multiple provinces and cities.

Other Models Worth Watching

The School of Earth Sciences at Zhejiang University released the “Fuxi Wanxiang Spatiotemporal Computing Foundation Model” in March 2025. Combined with the Zhejiang University “Mr. ZJU” agent and an Earth system big data platform, it performs geospatial-temporal computing and is currently more oriented toward teaching and research. Huawei Cloud's Pangu Weather Foundation Model appeared in Nature in 2023 and can produce global medium-range weather forecasts in seconds. The Institute of Geophysics, China Earthquake Administration, together with Tsinghua University and the National Supercomputing Center in Chengdu, developed “Diting”, the world's first hundred-million-parameter seismic wave foundation model, used for waveform recognition, magnitude estimation, and early warning. Besides SuperMap and GEOVIS, GeoStar, China Siwei, and AutoNavi have also integrated spatial agents or 3D generation into their products, but most have not yet been separately named as spatiotemporal foundation models, so they are not discussed further here.

Conclusion

In recent years, China has seen a flourishing of spatiotemporal foundation models. Some address how to interpret cross-sensor imagery, some enable models to read geographic literature and produce thematic maps, some are used for satellite image law enforcement, map review, and 3D reconstruction, and others focus on forestry and grassland patrol, earthquake emergency response, weather forecasting, and seismology.

Which models are you currently using? How well do they perform? Comments are welcome.

References

  1. National Surveying and Mapping Law Publicity Day | Comment: Promoting Deep Integration of AI and Surveying and Mapping: https://www.csgpc.org/detail/27975.html
  2. The world's first ten-billion-parameter remote sensing interpretation foundation model released (Aerospace Information Research Institute): https://aircas.ac.cn/dtxw/kydt/202409/t20240923_7379115.html
  3. The world's first multimodal geographic science foundation model “Kunyuan” led by the Institute of Geographic Sciences and Natural Resources Research released in Beijing: https://igsnrr.cas.cn/news/picnews/202409/t20240920_7374237.html
  4. Foundation models push geographic science toward geographic intelligence (Outlook): https://lw.xinhuanet.com/20241104/375c13ae9ad2413dab97a0fdbe7e58dc/c.html
  5. Two “Kunyuan” foundation models released: https://www.cas.cn/syky/202508/t20250807_5079293.shtml
  6. Mogan Xuanyan geospatial foundation model: https://www.mgslab.ac.cn/result/detail/id/10000.html
  7. Mogan Mountain Geospatial Information Laboratory releases two important scientific and technological achievements: https://www.mgslab.ac.cn/result/detail/id/10005.html
  8. China Natural Resources News | Multiple achievements released at the spatiotemporal geographic information pilot test base: https://www.mgslab.ac.cn/result/detail/id/10006.html
  9. Chinese Academy of Surveying and Mapping holds 2025 scientific and technological achievements promotion meeting: https://www.csgpc.org/detail/26423.html
  10. New prospects for spatiotemporal intelligence, new engine for industrial development: https://www.csgpc.org/detail/27131.html
  11. Zhongce Kunyuan foundation model (TerraMaster) photovoltaic extraction engineering production test: https://www.csgpc.org/detail/27412.html
  12. China's first forestry and grassland industry foundation model unveiled: https://www.forestry.gov.cn/c/www/zhzs/621257.jhtml
  13. Integrated intelligent foundation model for earthquake disaster prevention, reduction, and relief “Panyu” officially released: https://www.csgpc.org/detail/27915.html
  14. Sichuan integrated intelligent foundation model for earthquake disaster prevention, reduction, and relief “Panyu” officially released: https://www.chinanews.com.cn/sh/2026/08-18/10679934.shtml
  15. Xingtong foundation model product page: https://www.geovis.com.cn/products/detail/26
  16. GEOVIS Earth Intelligence Engine: https://www.geovis.com.cn/products/detail/20
  17. SkySense jointly launched by Wuhan University and Ant Group officially released: https://news.whu.edu.cn/info/1015/462657.htm
  18. DAMO Academy's first remote sensing AI foundation model released: https://www.qbitai.com/2023/10/91623.html
  19. AI Earth Geoscience Cloud Platform: https://engine-aiearth.aliyun.com/
  20. “Fuxi Wanxiang Spatiotemporal Computing Foundation Model” released: https://www.taibo.cn/p/97110
  21. Huawei Cloud Pangu Weather Foundation Model research published in Nature: https://www.huawei.com/cn/news/2023/7/pangu-ai-model-nature-publish
  22. China releases the world's first hundred-million-parameter seismic wave foundation model: https://www.tsinghua.edu.cn/info/1182/112989.htm
  23. “Earthquake Science Intelligent Agent” launched: https://epaper.gmw.cn/gmrb/html/content/202604/27/content_12154.htm