Shulou(Shulou.com)06/02 Report--

Just yesterday, the remote sensing image sparse representation and intelligent analysis competition sponsored by the Information Science Department of the National Natural Science Foundation of China and the major research project "basic Theory and key Technologies of Spatial Information Network" has just concluded. This is the third competition on sparse representation and intelligent analysis of remote sensing images organized by the National Natural Science Foundation of China. With the change of competition topics in the past three years, an obvious trend is that the combination of remote sensing and AI is becoming more and more abundant.

From the first year, intelligent image target detection and intelligent compression were set up, and in the second year, optical image recognition, SAR remote sensing image recognition and remote sensing satellite tracking were added. This year, in the final, the competition set up a remote sensing image interpretation plus score competition based on Huawei Penton AI processor, requiring the participating teams to transplant the final algorithm model to the Atlas 200DK AI platform to realize the reasoning calculation of the algorithm model, which actually shows the infinite opportunities created by the encounter between AI and remote sensing.

As a result, the winning team in this competition is particularly eye-catching-their results may bring more possibilities to the future of AI+ remote sensing.

Peak work: using a small number of samples to learn to make remote sensing image detection smarter

In the competition, the "smart team" composed of Deng Yuyi, Jie Yongshi, Zhang Yi, Chen Jing and Liu Wenya from the Aerospace Information Innovation Research Institute of the Chinese Academy of Sciences and Meng Yu won the special prize of this competition, and the topic they chose was remote sensing image change detection.

Remote sensing image change detection can detect and process different temporal data in the same location, which has a high practical value in the fields of resources and environment monitoring, geographical condition monitoring, natural disaster assessment and so on. At present, the cutting-edge algorithms of remote sensing image change detection have some limitations in the data level and algorithm design level, which leads to the inefficiency of the algorithm in practical application. The solution of the smart team is to apply the very hot small sample learning in AI for nearly two years. Under the premise that the data samples are not too abundant, we can improve the efficiency of data utilization through transfer learning and sample generalization as far as possible. In the end, the model of the witty team performed well in accuracy and efficiency, won the special prize of the whole game, and confirmed the infinite fit space between AI and remote sensing.

In fact, taking a closer look at the several unit settings of this competition, we can find that the industrial value of AI to remote sensing is being recognized by more and more people.

For a long time, the information captured by remote sensing technology has been applied to the research and observation in the fields of meteorology, environment, geological resources, agriculture and forestry and so on. Actual remote sensing as a technology that can detect targets over long distances beyond geographical obstacles, the accumulated image resources are valuable for many industries. However, most of the time, the analysis and observation of remote sensing images still need the personal participation of experts, and the manpower problem has become the key that remote sensing images can not be used by more industries.

At this time, through AI technology instead of manpower, remote sensing image data can be directly transformed into methodological tools that can be directly applied to the industry, which naturally becomes the key to get through the value of remote sensing technology industry.

For example, the problem of remote sensing satellite video tracking is to track vehicles such as cars, planes, trains, ships and other vehicles. This technology has great application value in intelligent logistics, intelligent transportation and other fields. In the semantic segmentation competition of remote sensing images, the organizers provide image data containing 15 typical land use types, and the analysis results of these data also have high application value for urban planning and construction in smart cities. The smart team that won the special prize can also help remote sensing technology to play a better role in accurate analysis and even prediction of disasters if they are applied to areas such as AI disaster prevention.

The key trick: how to make AI+ remote sensing over three hillsides by calculation

We noticed that in the final reasoning session, the smart team divided the computing task into two parts, one set in the cloud and the other applied edge computing. This kind of distributed computing greatly improves the reasoning efficiency of the model. Behind it is the Atlas 200DK AI developer suite provided by Huawei for the competition. This suite releases the strong computing power of Huawei Ascend 310chip through the peripheral interface, and has the characteristics of fast construction and easy iteration, which makes it convenient for participating developers to get familiar with it as soon as possible, and can quickly update and iterate the algorithm while constantly adjusting the strategy, so that participants do not have to waste time on the docking between the model and hardware.

In this case, we can also find that computing power is playing a more and more important role in the industrial logic of the combination of AI and remote sensing.

For example, where the computing power is, AI+ remote sensing is there.

Processing remote sensing images itself means the emergence of massive computing requirements. Whether it can meet different computing needs also means that AI+ remote sensing can sneak into more scenes. For example, in the field of low-and medium-altitude remote sensing, there are many tracking tasks, such as tracking the spread of disasters or tracking wildlife. This kind of time-tight and heavy computing work tests whether the UAV, cameras and other terminal devices can carry AI algorithm to achieve real-time identification and tracking, and also tests the computing ability of the terminal.

For example, how inclusive the computing power is, the more inclusive AI+ remote sensing is.

In the era of AI, we can deeply feel that computing power is like a currency, and the cost of completing AI tasks is clearly priced. However, from the industrial level, we must find a balance and inflection point between the costs and benefits of introducing AI+ remote sensing in order to dig out the application value of AI+ remote sensing. In other words, only by continuously lowering the threshold and cost of applying computing power can more industries apply AI+ remote sensing.

Finally, how perfect the computing ecology is, how perfect the ecology of AI+ remote sensing is.

When AI enables remote sensing to help remote sensing into industrial application, it also means that remote sensing image will become one of the many data processed by enterprises. If remote sensing image data want to be integrated into the overall technical architecture, it is necessary to enable remote sensing data to cooperate with the overall business of the enterprise in storage, transmission and processing. And this often tests whether enterprises have a perfect computing ecology to distribute and deal with different types of data, so that different computing models cooperate with each other. In other words, whether the application of AI+ remote sensing can be normalized depends on whether the computing ecology is perfect enough.

When the computing power helps AI+ remote sensing to climb over these three slopes, we will certainly see the dazzling light of remote sensing technology in more industries.

Leading climber Atlas: full scene AI and remote sensing meet at the peak

Going back to the event itself, we can see that many teams have used the Atlas 200 DK AI developer suite provided by Huawei to solve the computing needs in AI+ remote sensing. So can it provide more drivers for AI+ remote sensing?

The answer is yes. Atlas plays a full-scene AI, which is a full-scene AI infrastructure solution designed for "end, edge, cloud" optimization. The combination of AI technology and remote sensing, on the one hand, means massive computing requirements, on the other hand, in the application of target tracking, AI algorithm is also needed to respond to the data in real time, and there are also high requirements for end-to-side computing. In particular, remote sensing technology is often used in agriculture, meteorology, water conservancy and other "big scenes", naturally there is also the need for cloud cooperation.

Through the collation of Atlas products, we can sum up three kinds of industrial dividends released by full-scene AI to help remote sensing:

The first is the dividend of scene diversification.

The powerful end-to-side computing power brought by the Atlas 200AI acceleration module enables hardware often used in remote sensing fields such as UAV and embedded cameras to obtain real-time computing power to analyze and track remote sensing data directly on the end-side. Coupled with the affordable price of Atlas series products, it will not bring too much cost pressure to the industrial end. From drones photographing cities and farmland to satellite cameras, they can gain the ability to process computing tasks in real time, so AI+ remote sensing can also be quickly popularized in more scenes.

The second is the dividend of past data mining.

Since the 1980s, China's remote sensing technology has been active. But the combination of AI and remote sensing began in recent years. In other words, the huge amount of data accumulated in the past are all available treasures. The recently launched "strongest on the surface" AI training cluster Atlas 900 provides powerful computing power to efficiently mine past data and further promote the utilization of remote sensing technology.

Finally, there is a Yunbian synergy dividend.

Huawei has not only Atlas 200 AI acceleration modules that provide end-to-side computing power, but also Atlas 300 AI acceleration cards and Atlas 500 smart stations that provide edge computing power. The core competence of these products naturally comes from Huawei's Teng AI processor. The Teng system provides a full stack solution, including chip enabling, training and reasoning framework, and application enabling, as well as a full-scene deployment environment, including cloud, edge, Internet of things industry terminals and consumer terminals. The channel formed in this way can also enable AI remote sensing to achieve high expansion of business through cloud edge cooperation, reduce latency, and facilitate future development, so as to further meet the needs of the industry.

Under the several dividends brought by Atlas, AI and remote sensing achieve a leapfrog combination: AI not only accelerates the application of remote sensing in scientific research, but also helps remote sensing to move forward on the road of industrialization and commercialization.

Concluding remarks

It is not difficult to find that under the framework of the whole scene, the combination of AI and remote sensing meets like a peak. The application of remote sensing technology complements an important part of the image capture of the physical world-from tiny faces to nebula works, which can be transformed into image data for computer understanding and analysis. The full-scene AI provides an opportunity for computers to understand remote sensing images more efficiently, in drones chasing cherished wildlife at high altitude, and in servers that analyze climate change through meteorological remote sensing images for decades.

And calculation may be the climbing rope that draws the two to the summit, making it possible for AI and remote sensing to pursue a common goal.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.