Remote sensing meets the value and demand of AI, new opportunities and new challenges

As a universal technology that can penetrate thousands of industries, AI can often make a technology that sounds "old" instantly glow with new imagination. For example, remote sensing is a technical term that is not unfamiliar to Chinese people.

Remote sensing generally refers to the use of remote sensors to detect the radiation and reflection characteristics of electromagnetic waves of objects. The detection target is judged and identified by being far away from the target and non-contact. This technology is generally used in aerial platforms, such as satellites, aircraft, drones, etc.

100 years ago, the forerunner of remote sensing science was born in modern geography and surveying. In 1972, NASA launched the Earth Resources Technology Satellite ERTS-1 equipped with remote sensors, announcing the official arrival of modern remote sensing technology.

This technology, which has helped humans understand the earth for decades, is now having an imaginative encounter with AI technology. However, the combination of the two is not so easy. Breakthroughs in AI computing power on the edge and deep learning of algorithm models are becoming an indispensable cornerstone of the AI+ remote sensing industry and even the rising field of space intelligence.

For example, in June this year, the State Key Laboratory of Surveying, Mapping and Remote Sensing Information Engineering of Wuhan University and Huawei jointly held a competition on remote sensing image sparse representation and intelligent analysis. In this competition, which aims to promote the development of related theories and technologies for "sparse representation and fusion processing of spatial information", contestants will use Huawei's new intelligent computing product Atlas 200 DK development kit to implement inference calculations of algorithm models on the corresponding remote sensing image test data sets.

Today we can take this opportunity to explore the value and needs of remote sensing encountering AI.

When remote sensing meets AI, new opportunities and challenges

Since the birth of modern remote sensing, China has been a core participant in remote sensing technology with the help of the two bombs and one satellite project. From the 1970s to the present, my country has developed a complete range of remote sensing-related disciplines and diversified application fields. In the fields of environmental remote sensing, atmospheric remote sensing, resource remote sensing, marine remote sensing, geological remote sensing, agricultural remote sensing, forestry remote sensing, etc., there are a lot of application practices and cutting-edge explorations.

Today, the main opportunities and challenges facing remote sensing technology are to make this technology universal and accessible to all walks of life, and even to allow remote sensing to enter the vertical production cycle of industry, agriculture, and forestry, and become a "front-line worker" in the fields of urban planning, disaster prevention and relief, etc. There are many technical difficulties that remote sensing science itself cannot overcome independently. For example, in the field of drone remote sensing, the long cycle of image recognition and the large amount of time and labor costs required from data collection to data application have become one of the main constraints for remote sensing to enter various industries.

And this problem can be solved by AI.

We know that this round of AI technology represented by deep learning has brought an important capability, which is the machine vision technology system. Among them, image recognition, image processing, dynamic recognition, image classification and other capabilities can be applied to remote sensing data. Under ideal conditions, automatic recognition and reasoning from data to effective conclusion information can be achieved.

At present, many AI companies and research institutions in China have participated in the integration of AI + remote sensing industry, and have completed a large number of high-quality algorithm development in the fields of automated image processing and remote sensing data interpretation. Last year, China's first remote sensing artificial intelligence application technology research center was established in Chongqing. At present, my country has launched the exploration layout of remote sensing + AI in the fields of agriculture, industry, road network, meteorology, water conservancy, construction, etc.

Overall, the industry’s current understanding of “intelligent remote sensing” is that AI technology can provide active, real-time, and automatic error-correcting image recognition and reasoning capabilities in the remote sensing field.

Specifically for drone remote sensing, which is closely related to vertical applications in various industries, AI can bring the following help:

1. Complete a large amount of automatic identification work and realize the automatic processing of remote sensing data into usable data.

2. Shorten the remote sensing data usage process, and through automatic identification and automatic preprocessing, allow remote sensing results to be read and used in real time, thus making real-time remote sensing + operation linkage possible.

3. Use algorithms to restore image data and reduce the impact of the environment and weather. UAV remote sensing faces influencing factors such as complex terrain, rainy and foggy weather, etc., but these influencing factors can be partially eliminated through the restoration of specific algorithms.

4. Reduce labor consumption, improve surveying and mapping efficiency, and realize early warning remote sensing in some areas. For example, my country has adopted AI+remote sensing technology on plateau highway sections such as Yunnan-Tibet and Qinghai-Tibet to proactively detect mudslides and landslides.

However, Rome was not built in a day. Although the combination of remote sensing and AI can solve a large number of problems and build new imagination, in the actual combination, there will still be many industrial obstacles.

For example, as mentioned above, AI algorithms can improve the ability to automatically identify remote sensing images in real time, but in actual application, remote sensing satellites and drones need to upload data to the cloud, process it in the data center, and then transmit it back, which still cannot achieve real-time results. Uploading to the cloud limits the feasibility of applying AI remote sensing technology in the industry. For example, many industries related to national economy and people's livelihood are not suitable for large amounts of data to be uploaded to the cloud.

To this end, the best solution is to enable AI computing on end-side devices such as drones and remote sensors, pre-process data locally, and keep up with the production process in real time. However, the industry reality is that in the current AI+remote sensing industry, most companies are mainly solving algorithm problems. However, algorithm problems require an effective network environment and computing environment to ensure them. Just like the best home appliances, they are useless without electricity.

Therefore, we can see that in the remote sensing system, the hardware level, especially the supply of AI computing power on the edge side, becomes very important.

From another perspective, although AI reduces the manpower consumption required to identify remote sensing data, the complex hardware environment is likely to lead to more consumption of AI technical talents. In the current industrial environment, the cost of AI hardware talents is even greater. Therefore, ensuring the compatibility and environmental availability of AI capabilities in the remote sensing system on the hardware side is also an important issue.

Fortunately, edge-side AI computing solutions have already emerged in the industry. For example, Huawei's Atlas intelligent computing product has brought a surprising breakthrough to the remote sensing industry: allowing AI computing power to fly in the air.

Flying AI computing power

In the above logic, we described the current situation: if remote sensing wants to enter all walks of life, it is necessary to strengthen the applicability of drone remote sensing. The core issue is that the drone itself must have relatively sufficient AI computing power, so that image processing, image recognition, environmental recognition and other related operations can be directly performed in the device. This avoids the security issues that may be caused by uploading data to the cloud, while also speeding up the processing speed and shortening the business process.

In short, the most effective solution to this problem at present is to increase the AI ​​computing power and install AI chips on the drone itself.

It is not easy to achieve AI acceleration capabilities while flying in the sky. This requires sufficient AI computing power, multi-link visual data processing capabilities, and environmental adaptability in drone scenarios for collective assurance.

Back to the competition held by the State Key Laboratory of Surveying, Mapping and Remote Sensing Information Engineering of Wuhan University, contestants will use the Atlas 200 DK AI developer kit to complete the inference calculation of the data algorithm model. Imagine if not only the algorithm side introduces the Huawei Atlas intelligent computing platform, but also the drone equipment is equipped with the Huawei Atlas 200 acceleration module (Atlas 200 AI acceleration module is an embedded AI module, which is mainly used for end-side AI acceleration of hardware such as cameras, drones, and robots. In the field of video analysis, it can process real-time analysis of 16 channels of high-definition video. The matching Atlas 200 DK AI developer kit can complete the development environment construction within 30 minutes and provide up to 16TOPS INT8 computing performance.) Remote sensing will be empowered by powerful AI computing power on both the end measurement and algorithm sides, which will greatly improve the work efficiency in remote sensing mapping.