Cheap commercial space is the key
The author: Remote Sensing Division of Science and Technology Shang Zhang Lin, Dr. summer
Original title: “Imagine New Technology for the Next Decade-Remote Sensing AI Interpretation Technology”
Editor: Shi Yaqiong
Any kind of disruptive new technology from vision to mature application, from “spark of thought” to “finished material” should have three of the most prominent characteristics: 1. It can effectively solve at least one of the human activities Questions; 2. It is universal and reusable within a certain range; 3. It has economic value or social benefits.
Remote sensing technology was born in the 1960s. After decades of rapid development, it has become a practical and advanced space detection technology, but still faces many constraints. In the next ten years, whether remote sensing can profoundly affect social development, practically solve problems in production and life, and have universality and economic value. The key points are the interpretation and application of remote sensing data, artificial intelligence technology and remote sensing. Combination may be a golden key that opens the door to future applications in the remote sensing industry.
Remote sensing and artificial intelligence
Remote sensing, literally, can be simply understood as distant perception, which refers to all contactless long-range detection. Traditional remote sensing interpretation technology is not ideal for accurate and fast processing, and lacks effective means for refined state analysis. . The most restrictive is that the image interpretation method mainly relies on manual interpretation and semi-automatic software interpretation, which makes remote sensing applications unable to fundamentally break away from its labor-intensive “traditional”.
Since 2015, the number of earth observation satellites that have been launched and are operating in the world has increased significantly from 223 to 710, with the subsequent expansion of the satellite remote sensing data analysis market. According to the satellite consulting company NSR, by 2027, the total global satellite data analysis market will reach 18.1 billion US dollars.
The surge in the amount of multivariate remote sensing data, the huge prospects of the remote sensing data analysis market, and the bottlenecks of traditional remote sensing technology urgently require a new, efficient, accurate, and convenient technical method to fill it. Artificial intelligence is a branch of computer science. Since the 1970s, space technology and energy technology have been called the world’s three major cutting-edge technologies. Today, one of the cutting-edge applications of space technology ——The combination of remote sensing technology and artificial intelligence technology will enable artificial intelligence to enable remote sensing technology to run through the entire link of massive multi-source heterogeneous data from processing analysis to sharing applications, greatly reducing the interpretation cycle of remote sensing images and improving the accuracy of interpretation. At the same time, it has spawned new remote sensing application fields and promoted the transformation of remote sensing technology applications.
Artificial intelligence + remote sensing
With the rapid development and widespread application of artificial intelligence technology in recent years, and the urgent need for new interpretation capabilities of remote sensing technology, more and more high-tech companies and research institutions have begun to try to solve massive remote sensing images using deep learning. The problem of interpretation and some staged progress have been made and put into application in the remote sensing industry. The SenseEarth intelligent remote sensing online interpretation platform released by Shangtang Technology on WGDC this year and the SenseRemote intelligent remote sensing interpretation series products behind it.
However, although some progress has been made in the combination of artificial intelligence and remote sensing technology at this stage, the use of deep learning technology to interpret remote sensing image processing accuracy, efficiency, and degree of automation in some application scenarios is relatively obvious. To improve, we have to face the limitations of the current results and the huge challenges facing future development.
First, most of the current AI remote sensing applications adopt the method of supervised learning. The basis for intelligent interpretation of massive remote sensing data using such technologies is the pre-training of the same mass of labeled samples of specific interpreted objects. Work, which not only requires a large amount of computing resources and sophisticated design capabilities, but more importantly, the need for labeled samples of remote sensing data during the model building process. The application scenarios of remote sensing are extremely rich and diverse, and even the same interpretation object exhibits different characteristics in different spatial and temporal dimensions, which makes the complexity of data samples increase geometrically. As a result, it is currently impossible for an institution to change most The correctly labeled sample sets in the remote sensing application field are collected into a library to train an effective interpretation model.
Taking the agricultural remote sensing industry, which is leading in the development of remote sensing technology, as an example, the application includes dozens of specific application scenarios such as crop classification and identification, yield estimation, growth analysis, field soil moisture, and pest control, etc. In terms of crop classification and identification, there are more than 390,000 species of plants on the earth, including about 2,300 cultivated plants and about 90 species of crops used by humans. There are more than 60 common crops in China. The same type of crops are in different regions or even the same region The characteristics presented in different geographical environments are different, and this complexity makes it difficult for the remote sensing intelligent interpretation model obtained by deep learning based on supervised learning to be universal and reusable.
Second, the multiple isomerization of remote sensing data sources, different remote sensing platforms such as rotary wing drones, fixed wing drones, manned aircraft, near space airships, low orbit satellites, and high orbit satellites; imaging mechanisms for different loads such asVisible light, SAR, hyperspectral; different spatial-temporal spectral resolution, accuracy, timeliness, etc. all bring huge challenges to the consistent processing of remote sensing data. How to use multi-source heterogeneous data to build a “one picture” style The application scenarios that make artificial intelligence technology to easily and conveniently solve the problem of extracting and analyzing the spatiotemporal information of massive heterogeneous data will be the top priority of the development of the remote sensing industry.
Finally, in view of the comprehensive development of artificial intelligence remote sensing technology, its development depends not only on the technical iteration and development of remote sensing and artificial intelligence itself, but also on computer technology, storage technology, aerospace technology, Internet sharing technology, and even mathematics and neuroscience. The related technological and theoretical innovations in various fields will affect the advancement speed of the artificial intelligence remote sensing industry to a certain extent. This is similar to a large and complex system engineering. Any one of these links may become a limiting factor or a development aid. . This makes artificial intelligence + remote sensing technology have a long R & D cycle and capital risks before it can generate extensive economic benefits.
The future of artificial intelligence remote sensing
Sample accumulation
In view of the dependence of the artificial intelligence remote sensing interpretation deep learning algorithm model on the large number of labeled samples at this stage, the use of emerging network sharing technologies such as cloud and blockchain will be scattered in various industries and governments under a win-win mechanism. The remote sensing samples in institutions, research institutes, and companies are integrated and supplemented with each other. At the same time, using the development of data simulation technology to jointly build an interpretation model library belonging to a large industry category may be a way to solve the lack of samples in the development of intelligent remote sensing technology. one. But this requires a good business model, so that all parties involved can benefit while paying. It needs a healthy industrial ecosystem to allow data resources, computing resources, and scientific research resources to flow unhindered. It requires a The long-term layout and rules make this integration work healthily. This is undoubtedly very difficult.
Fortunately, we have already seen the efforts and efforts of related companies in the industry. For example, the planning of Shangtang Technology ’s SenseEarth platform mentioned that “In the future, a lightweight online sample training platform system will be launched on the basis of SenseEarth, hoping to have more communication and cooperation with users.
Unsupervised learning
From another perspective, the foundation of current deep learning is training on a large amount of structured sample data that is correctly labeled. Although we are in an era of information explosion and flooding with various data, most of these data, especially remote sensing data, are unlabeled and unorganized, which means that these data are not available for most current supervised learning. Yes, even though there are now a large number of free and public label datasets covering a wide variety of different categories, it turns out that they interpret artificial intelligenceThe role of remote sensing data is still very limited.
The labeled sample set may be too small, or the label may be biased. When training a complex remote sensing interpretation model, using a small data set may lead to so-called overfitting, which is due to the large number of learnable parameters and the training sample. As a result of the association, what we end up with may be a model that is only applicable to these training samples, rather than a model that learns general concepts from the data.
Unsupervised learning algorithm will be an important technical development direction to solve the scarcity of labeled samples in remote sensing data. Unlike supervised learning, which requires labeling and classification in advance, it can be very helpful to us based on the unknown and unlabeled training samples. Solve various problems in the interpretation of remote sensing data. In the face of massive remote sensing data, we need to deal with not only the various samples that are structured and labeled, but the remote sensing data itself. In simple terms, unsupervised learning enables the machine itself to perform clustering and analysis on image datasets. For example, there are many detectable targets such as cars, airplanes, ships, buildings, and roads in remote sensing images. The method of supervised learning is to first These targets are labeled, features are extracted for training, and then target recognition is performed based on the training results. Unsupervised learning is based on the characteristics of algorithm perspectives that humans may not understand, which are summarized by these targets. The feature comparison model recognizes a class of objects.
Decision-based artificial intelligence interpretation
In actual business scenarios, what we need to give is often a comprehensive solution, which means that the establishment of interpretation models must be based on the fusion of heterogeneous remote sensing data of different platforms, different loads, different phases and different scales. Multi-category targeted analysis methods work together to draw conclusions. In the past, artificial intelligence remote sensing was mostly a migration of traditional digital image processing methods, and even only solved the problem with statistical concepts. With the development of artificial intelligence technology, decision-making intelligent technology will become one of the mainstream development goals in the future. The decision here refers not only to the use of results to help users make judgments, but to let the system bring its own when intelligently interpreting data Decision function. Just like people’s learning and thinking, when analyzing problems, they use “experience” to choose their own judgment basis. What kind of data is used, what are not so typical but effective features, targeted adaptation to specific scenarios, including professional network model adaptation, autonomous construction of heterogeneous entity networks, dynamic optimization of many-to-many relationships, etc.
In the future, when we have established a sufficient number of discrete intelligent interpretation models for segmented target objects, we may need a method that can sum up a large number of model libraries. Based on the accumulated model design experience, the system can further modularize the model and establish a model search space. Through reinforcement learning,