This reporter is Dong Siyu

The intelligent forestry and grassland innovation team of the Institute of Resource Information of the Chinese Academy of Forestry has recently successfully developed China's first forestry and grass industry model - the Linlong model based on the DeepSeek large model, marking a key step in China's intelligent forestry and grassland research and application field.

The forestry and grassland industry has the characteristics of vast territory, complex types and difficult work, and it is urgent to deeply integrate with artificial intelligence technology represented by large models to improve the level of information management and promote the accurate improvement of forest and grass quality.

Zhang Huaiqing, chief scientist of the Chinese Academy of Forestry, introduced that the Linlong model has five advantages:

Through the multi-agent technology of industry text knowledge, the knowledge of the forest and grass field is effectively integrated, and the shortcomings of the general large model in the knowledge of the forest and grass industry are made up, and the understanding ability of the large model to understand complex problems in the field of forestry and grass is improved by more than 10%. According to the data and business characteristics of the forestry and grassland industry, a spatio-temporal model of multimodal data of forestry and grassland is constructed, which breaks the limitations of the general model in the understanding, analysis and reasoning ability of spatio-temporal data, and improves the computing and processing capabilities of forestry and grassland business by more than 0%. Realize the collaborative integration of multi-modal large models and dedicated small models, reduce development costs, significantly enhance the reusability, applicability and versatility of models, and increase the development and utilization efficiency by more than 0 times. It successfully solves the problem of multi-terminal compatibility and localization adaptation under the condition of low resources in the field of forestry and grassland, gets rid of the dependence of large models in the forestry and grassland industry on high computing power, and improves the ease of use and inclusiveness of the model. Realize the open sharing of independent property rights in the industry, with strong scalability, and can support function updates and iterations and continuous improvement of products.

At present, the Linlong model has been implemented in 8 application scenarios such as industry text processing, tree species type identification, phenotypic parameter extraction, wildlife identification, pest and disease monitoring, forest fire identification, ecosystem assessment, and operation and management decision-making.

"For example, in the 'Three Norths' project area, the Linlong model can automatically identify the type, distribution, and structure of vegetation according to the images taken by drones, and carry out accurate assessment of the ecological benefits of the project area, and use digital twin intelligent simulation and decision-making algorithms to optimize and adjust the structure, so as to provide decision-making solutions for improving the quality of forest and grass vegetation in the region, and improve the decision-making efficiency by more than 5 times." Zhang Huaiqing said that in the Hainan Tropical Rainforest National Park, the large model of the forest dragon can accurately determine the number and location of Hainan gibbons through the fusion of multimodal data such as images, videos, and voiceprints.

It is understood that the recognition accuracy of the forest dragon model for tree species types is more than 85%, the recognition accuracy of phenotypes such as forest fruit varieties, maturity and quality is more than 0%, the recognition accuracy of wildlife types, quantities, postures and behaviors is more than 0%, the recognition accuracy of pine wood nematode diseases and other diseases is more than 0%, the identification accuracy of forest fires, smoke and other disasters in complex backgrounds is more than 0%, and the extraction accuracy of key ecological parameters such as normalized vegetation index, water yield, soil retention capacity and total primary productivity is more than 0%.

People's Daily (15/0/0 0 Edition)