Artificial intelligence at the Earth final frontier: a quick overview on how neural networks help Arctic operations, exploration and research
As the Arctic becomes an important commercial and industrial center, issues of safe, cost effective and environmentally friendly work in the region are becoming increasingly important.
Vessels moving in ice-covered waters along the Northern Sea Route should constantly monitor the ice situation both in the immediate vicinity and along the entire route, as well as the general regional ice and weather conditions to ensure the safety of the vessel, cargo and crew. in addition, oil and gas exploration and extraction also carry enormous risks for both oil-and-gas companies and people and the environment.
Now, scientists and engineers, one way or another connected with the study of the Arctic and the development of solutions to ensure its safe, economical and environmentally friendly development, face a whole series of tasks.
This includes ice edge detection to understand the dynamics of the ice mass, and the classification of the ice cover in terms of its age and thickness to ensure the safety and cost effectiveness of shipping convoys and the organization of the work of such unique projects as the “Prirazlomnaya” oil rig, and the determination of ice concentration and dispersion zones, which together with the prediction of drift makes it possible to assess in advance how the situation will develop and ensure safety, as well as the economic and environmental efficiency of operations beyond the polar circle.
All these tasks have one thing in common — their solution requires tremendous human effort, a long time and a large base of expert knowledge. All these factors negatively affect both the speed of development of the Arctic and the economics of operations. Sometimes experts issue their forecasts or classification results too late, when the actual situation has already changed, or, on the contrary, ships are waiting for reports in ports, wasting valuable time.
However, due to the recent rapid development of Artificial Intelligence technologies and Big Data, primarily with such a domain as Computer Vision, it became possible to automate the solution of the above tasks, thereby speeding up the decision-making process and taking experts block of laborious tasks.
One of the main sources of data in Arctic research and development is the so-called Remote Sensing of the Earth, or, in short, RS.
RS data includes video and photographs from drones, as well as the results of optical and radar scans obtained from satellites.
Currently, most of this data is processed manually — experts pore over the images obtained by hand, highlighting certain areas of the ice cover, determine areas safe for navigation and calculate the possible consequences of ice drift. Sometimes it takes days or even weeks to solve these problems.
At MaritimeAI we have developed Computer vision-driven methods that can reduce the time it takes to complete these tasks to hours or minutes, so ship captains and navigation officers can get analysis results online, saving valuable time when making safe, economical and environmentally friendly decisions.
From the point of view of technology, it looks like this: first a data set is collected with images from UAVs or from satellites, in one or more feed ranges (optical, radar, infrared). The more you collect this data, the more accurate the algorithm will work in the future — the golden rule of machine learning works: more good data and an average algorithm give a better result than little data or bad data and a very good algorithm.
Once the dataset is assembled, the time comes for the most time-consuming task — markup. At first, the data is divided into 3 different sets — training, i.e. the one on which the algorithm will learn, the test one with which the algorithm will be checked, and the validation one, from which the algorithm has never seen images and on which it will try to predict the necessary classes.
Experts, as usual, manually mark up the data, i.e. the necessary areas with different thicknesses of ice or, for example, ice with different ages are highlighted on the images. The more such marked images, the better. Then, specialists in computer vision build an algorithm, select the optimal parameters and begin training.
Training is considered successful when the algorithm has learned to perform the task with the necessary accuracy. After that, experts on the validation dataset evaluate the results and give their verdict — accept the result or not. If the result is accepted, then the algorithm passes from the training stage to the stage of the so-called “inference”, i.e. no longer studying, but begins to automatically determine the topology of the ice situation. The results are then used to draw ice maps in real time or to build optimal routes.
Of course, there is no such thing as a fee lunch. For example, working with images from radars is much harder than with an optical channel and the initial markup time will be significant, in comparison with all other stages of work.
In addition, to improve the quality of the predictive model, sufficient sets of good data are required, in 2–3 or even 10 images the neural network is poorly trained. Nevertheless, the effect obtained by AI in the tasks of developing the Arctic far exceeds the difficulties described above and allows us to believe that more and more organizations operating in the Arctic Circle will turn to neural networks for help and humanity will approach the dream of the Arctic as an ally, not a stern adversary.
