When Traditional Remote Sensing Meets Artificial Intelligence: What Does the Future Hold?

2019-12-27

In recent years, with the development of smart city construction, new urbanization planning, and the refinement of agricultural and forestry resource statistics, the commercialization of remote sensing imagery has been greatly promoted. However, the available remote sensing data in the market, in terms of both quality and quantity, still lags behind the growing demand. The emergence of artificial intelligence technology comes at the right time. Any disruptive new technology, from vision to mature application, from "spark of thought" to "material product," should have three most significant characteristics: 1) it can effectively solve at least one type of problem in human activities; 2) it has universality and reusability within a certain range; 3) it has economic value or social benefits. Remote sensing technology was born in the 1960s and has rapidly developed over the past few decades, becoming a practical and advanced space exploration technology. However, it still faces many constraints. Whether remote sensing can profoundly impact social development, effectively solve problems in production and life, and have universality and economic value in the next decade depends on the interpretation and application of remote sensing data. The combination of artificial intelligence technology and remote sensing may be the key to opening the door to future applications in the remote sensing industry. Remote sensing, as the name suggests, can be simply understood as remote perception. It refers to all non-contact long-distance detection. Traditional remote sensing interpretation techniques have unsatisfactory results in precise and rapid processing and lack effective means for fine-grained state analysis. The most limiting factor is that image interpretation methods mainly rely on manual reading and semi-automated software interpretation, which prevents remote sensing applications from fundamentally breaking away from their labor-intensive "tradition." From 2015 to the present, the number of working Earth observation satellites launched globally has increased from 223 to 710, leading to the expansion of the satellite remote sensing data analysis market. According to satellite consulting firm NSR, the global satellite data analysis market is expected to reach a total of $18.1 billion by 2027. The surge in diverse remote sensing data, the huge prospects of the remote sensing data analysis market, and the bottleneck of traditional remote sensing technology urgently require a new, efficient, accurate, and convenient technical means to bridge the gap. Artificial intelligence is a branch of computer science and has been considered one of the world's three cutting-edge technologies along with space technology and energy technology since the 1970s. Nowadays, the combination of remote sensing technology, one of the frontier applications of space technology, and artificial intelligence technology empowers remote sensing technology, spanning the entire chain from processing and analysis to shared applications of massive multi-source heterogeneous data. This combination significantly shortens the remote sensing image interpretation cycle, improves interpretation accuracy, and gives birth to new remote sensing application areas, promoting the transformation of remote sensing technology applications. Traditional remote sensing image interpretation methods mainly rely on manual reading and semi-automated software interpretation. The former involves professional personnel conducting map interpretation and analysis, which is relatively inefficient and has a high entry barrier. The latter uses software to quantitatively analyze images, but its accuracy and efficiency improvement have reached a bottleneck, and it still depends on the professional knowledge level of end users. With continuous breakthroughs in algorithms and computing power, artificial intelligence has broken the labor-intensive "tradition" of remote sensing applications. Currently, in several implemented intelligent remote sensing interpretation overall solution projects in various regions, SenseTime can provide product services for the entire ecological chain, including data acquisition, basic platform construction, algorithm interpretation library construction, and thematic applications. Considering the current stage, the construction of deep learning algorithm models for artificial intelligence remote sensing interpretation relies on a large number of annotated samples. By utilizing emerging network sharing technologies such as cloud and blockchain in a win-win mechanism, scattered remote sensing samples in government agencies, research institutes, and companies in various industries can be integrated and complement each other. Meanwhile, with the development of data simulation technology, a comprehensive interpretation model library belonging to the large industry category may be one of the ways to solve the problem of insufficient samples in the development of intelligent remote sensing technology. However, this requires a sound business model that allows all parties involved to benefit from their contributions. It also requires a healthy industrial ecosystem that allows the free flow of data resources, computing resources, and research resources. It needs a long-term layout and rules to ensure the healthy operation of this integration. Undoubtedly, this is very challenging. Actual business scenarios often require a comprehensive solution, which means that the establishment of interpretation models must be based on the fusion of different platforms, payloads, time phases, and scales of heterogeneous remote sensing data. Various targeted analysis methods should work together to draw conclusions. In the past, most artificial intelligence remote sensing mainly relied on the migration of traditional digital image processing methods, or even only used statistical concepts to solve problems. With the development of artificial intelligence technology, decision-making intelligent technology will become one of the mainstream development goals in the future. Here, decision-making does not only refer to using results to help users make judgments but also means that the system has decision-making capabilities when interpreting intelligent data. Just like human learning and thinking, when analyzing problems, it autonomously selects judgment criteria based on "experience." It adapts to specific scenarios by using different data types, less typical but effective features, and includes the adaptation of professional network models, autonomous construction of heterogeneous entity networks, and dynamic optimization of multiple associations. In the future, when enough intelligent interpretation models for subdivided target objects are established, a method that can summarize and summarize a large number of model libraries may be needed. A system that can achieve self-learning iteration and self-decision-making, based on accumulated model design experience, can further modularize models and establish a model search space. Through reinforcement learning, it can search for targeted models that are more suitable for its own problems within the search space.

RECENT AI NEWS

RECENT AI TOOLS