Going deep into thousands of industries, what can AI remote sensing do?
Author | Sun Jia
Editor丨Gao Yan
Source | Mustang Finance
Since the iterative wind caused by the artificial intelligence big model has blown from abroad to China, all major manufacturers are rubbing their hands, and related products have sprung up. As one of the leading manufacturers, SenseTime (0020. HK) naturally attracts the attention of the industry.
Recently, WGDC (Global Geographic Information Developers Conference) 2023, known as the "scientific and technological innovation vane" of spatiotemporal information, opened in Beijing. At the conference, SenseTime displayed the remote sensing large model built under the large model system, and released the SenseEarth 3.0 intelligent remote sensing cloud platform and the "SenseTime" smart remote sensing data product.
Remote sensing technology is known as the "eye of the sky". In short, with the help of artificial earth satellites, this technical means can interpret and analyze "photos" at a distance, and freeze the shape and distribution of various things on the earth. The combination of AI and remote sensing can greatly improve the processing efficiency of remote sensing image interpretation, so as to output more refined, more accurate and more beautiful extraction and interpretation results. On this basis, AI remote sensing can complete a large number of automatic identification work in agriculture, green finance, natural resources, environmental protection, urban governance and other fields, shorten the data use process, reduce manual consumption, and better help the industry reduce costs and increase efficiency.
However, the application of AI remote sensing has long had problems such as high threshold, high cost, multiple heterogeneity of data sources, and difficulty in interpretation, which has become a constraint for this technology to produce extensive economic benefits.
In this context, SenseTime's SenseEarth 3.0 intelligent remote sensing cloud platform and "SenseTime" smart remote sensing data products can not only provide better services for the traditional remote sensing application market, but also greatly reduce the threshold for the use of intelligent remote sensing. As a result, AI remote sensing can sink into various industry segmentation scenarios.
When all walks of life have a "God's perspective", how will the world change?
The advantages of large models have made a new chapter in remote sensing applications
"After a major breakthrough in deep learning, AI is already at an inflection point from 1.0 to 2.0." This is the judgment of Kai-Fu Lee, chairman and CEO of Innovation Works, in March this year.
In Kai-Fu Lee's view, AI 2.0 can overcome the limitations of the previous single-field and multi-model, and AIGC is the first phenomenal application in the 2.0 era. Taking ChatGPT as an example, its predecessor, GPT-3 model, used more than 175 billion parameters, trained on about 45TB of data, contained nearly 1 trillion words of text, and was equivalent to about 13.51 million Oxford dictionaries.
These figures are the verification of the mainstream route of large models in the field of AI, which is the field that SenseTime has accumulated for many years.
As early as around 2018, SenseTime has begun to prepare for the computing power, algorithms, and data required in the early stage of the big model. At the beginning of 2022, SenseTime AIDC officially launched operations, AIDC is an important computing power base for SenseCore's SenseTime AI device, and is currently one of the largest artificial intelligence computing centers in Asia.
On this basis, SenseTime's large model has made many breakthroughs in many fields such as visual perception, language understanding, content generation, and reasoning decision-making.
Compared with other large models at home and abroad, SenseTime's full-stack large model R&D system has made core technological breakthroughs in visual large models by virtue of the advantages of multimodal large models.
SenseTime's AI remote sensing model is based on its general vision model. With the foundation of general capabilities, enterprises or developers only need to fine-tune on the basis of pre-trained models to "produce" applications that are competent for specific scene tasks, just like industrialized mass production of goods, so that AI remote sensing can achieve simple and efficient development.
It is foreseeable that AI remote sensing has a wide range of application prospects, such as land cover analysis, terrain analysis, urban planning, agricultural monitoring, natural resource management and other fields. The lower cost and more efficient performance it brings will also help the upgrade and promotion of digital technology in all walks of life.
The economic value of large models of remote sensing
For a long time, as a data-intensive service, remote sensing has been suffering from problems such as shortage of computing power and inability to release the value of data. A standard image map can often reach billions of pixels and cover tens of thousands of square kilometers, which is difficult to interpret with only previous deep neural network models. This is also the significance of SenseTime's construction of AI remote sensing large model.
By sharing structure and parameters with the general vision model, SenseTime's AI remote sensing model has achieved breakthroughs in interpretation accuracy and time.
Specifically, SenseTime's AI remote sensing model has achieved high generalization capabilities for different species, different image types, and different image times across the country, has advanced feature interpretation capabilities and generative patch effects comparable to artificial annotations, and covers 46 types of semantic segmentation, 5 types of target monitoring, 4 types of change detection, and 2 types of super-resolution algorithms.
In terms of efficiency, in the past, it took 25 units of time to process 25 different semantic segmentation models, but now the SenseTime AI remote sensing large model only needs 1 unit time. In terms of accuracy, the average accuracy of its feature segmentation capability on the million-level patch verification set exceeds 80%, which can directly meet the application requirements of various business scenarios.
For example, SenseTime's cultivated land detection model has a plot identification effect comparable to manual sketching, which can complete the identification of 2,000 square kilometers of cultivated land in the southern region in 8 hours, and the work efficiency is more than 60 times higher than that of manual operation; SenseTime's crop detection model can complete crop identification in a district and county within 5 hours, and the work efficiency is increased by 5 times compared with manual work, and the larger the operation range, the more significant the efficiency.
It is worth noting that the change in AI remote sensing does not stop at the fantastical numbers. Previously, although satellite companies also held a large amount of data, due to insufficient data processing capabilities, they usually could only face government customers, and the overall monetization ability was weak. However, after service providers represented by SenseTime revitalized data resources through AI technology platforms and made them data products that can also be consumed by the general public, the industrial chain has been shortened, and the value of data elements itself is constantly expanding.
At present, based on the AI remote sensing model, SenseTime's remote sensing business has served more than 20,000 industry users, covering natural resources, agriculture, finance, environmental protection, photovoltaic and other industries.
Especially in the field of natural resources, which is known as "the first of the 100 industries of remote sensing", SenseTime's AI universal change detection has been widely used in natural resource law enforcement supervision in more than 14 provinces and cities, including Zhejiang, Guangdong, Guangxi, Yunnan, Guizhou, etc., with its capability advantages, helping users improve their work efficiency by 3-5 times.
In addition, SenseTime's AI remote sensing model has also been applied in the fields of non-agricultural and non-food monitoring, food security monitoring, photovoltaic roof census, green finance, and bare ground dust.
It is not difficult to see that the AI remote sensing large model has expanded into rich application fields based on its accurate identification of target scenes. Under this pair of "eyes of heaven", everything from a small crop to a large arable land and a city can be seen at a glance. With the help of this tool, farmers can control food production, merchants can screen suitable location resources, and managers can grasp the construction of cities... As long as there are scenarios with geographic information, AI remote sensing can cover and provide high-quality interpretation services to help users in various industries reduce costs and increase efficiency.
When remote sensing is no longer far away
After improving the recognition accuracy of information, how to reduce the user's threshold for use has become the focus of the industry's next consideration. After all, even if there are good products, if they are not favored by users, they cannot form a commercial closed loop.
However, due to the abundance and diversity of remote sensing applications and the heterogeneity of data sources, this is not an easy task.
In this regard, SenseTime's launch of the "Data-as-a-Service" data product DaaS (Data-as-a-Service) service with the SenseEarth 3.0 platform may bring new enlightenment to the industry.
Specifically, in addition to providing traditional remote sensing images, SenseTime also provides a complete set of structured vector data, which can intuitively see the situation of various features, which can be directly used for analysis, help users solve business problems, and greatly reduce the application threshold of intelligent remote sensing.
At the same time, derived from the powerful analysis capabilities of AI remote sensing large models, the SenseEarth 3.0 platform also provides more than 30 intelligent remote sensing analysis algorithms, covering application scenarios such as feature classification, change detection, and target recognition, and has rich application components such as maps, charts, text, and timelines, supporting visual and free layout. With the help of these tools, users can realize the construction of different business systems such as natural resources, agriculture, and new energy.
At present, the platform has launched more than 120+ "SenseTime" data products, with a total area of 87,146 square kilometers. In addition, SenseTime is expected to release the "SenseTime" data product covering the whole country by the end of this year.
It is worth noting that with the continuous improvement of space infrastructure construction, the geographic information industry is becoming a new growth pole of the domestic digital economy. In this context, the satellite remote sensing service industry has ushered in a series of policy support since 2014.
For example, in June 2022, the State Council mentioned in the "14th Five-Year Plan" National Emergency System Planning Notice that it is necessary to steadily promote the construction of satellite remote sensing networks; In February this year, the State Council clearly stated in the "Overall Layout Plan for the Construction of Digital China" that it is necessary to accelerate the construction of a smart and efficient ecological environment information system.
The excellent policy environment coupled with the steady growth of the expenditure of agriculture, forestry and water subjects of government departments has laid a solid foundation for the development of the satellite remote sensing service industry.
What's more, as big models penetrate the industry, they will become more "smart". Huajin Securities pointed out that in the long run, large models represented by GPT continue to emerge new capabilities with the training of massive data, and remote sensing data naturally has the characteristics of regular standards, high quality, and continuous accumulation, and is expected to form new value after AI training.
For SenseTime, after connecting more customers and partners, its model will produce higher quality data, which will be fed back into the model after the cycle, so that the model can be iterated again, forming a virtuous circle.
Every leap in the level of intelligent technology has brought great progress in social productivity, and now AI remote sensing has shown the momentum of "flying into the homes of ordinary people". It is conceivable that in the future, it will play an important role in benefiting urban construction and social livelihood, and promoting the transformation and upgrading of economic structure.
What is your understanding of remote sensing? Can my work be combined with AI remote sensing? Welcome to leave a message in the comment area to discuss.