Artificial Intelligence for Effective Realization of Underwater Domain Awareness Framework

Underwater Domain Awareness and Computational Advancement

"Acoustic Capacity Building has a wide array of military and commercial applications concerning with Maritime Security, Blue Economy, Marine Environment, Disaster Management, Underwater Archaeology and many more"

The Underwater Domain Awareness (UDA) framework, proposed by the Maritime Research Center (MRC), Pune, essentially focuses on effective realization of the Digital Ocean initiative for enhanced maritime governance. This will facilitate enhanced transparency with improved sonar deployment in terms of range, robustness to medium fluctuations, data integrity etc. This is commonly referred to as Acoustic Capacity Building effort and has a wide array of military and commercial applications concerning with Maritime Security, Blue Economy, Marine Environment, Disaster Management, Underwater Archaeology and many more. Although the framework has been coined in recent years, underwater research works contributing towards enhancing performance of marine systems has been a continuous effort since the Cold War period – which required intensive study of propagation of sound in sea for Anti-Submarine Warfare (ASW) operations. A fine example of this would be the ‘Ross Model’, developed by Donald Ross (acoustician), who used the noise measured by SOSUS in World War II (WWII) and during the Cold War to develop a mathematical model that predicts the noise emitted by ships using its length and speed.

Over the years, technology has helped bridge the gap between theoretical assumptions and practical implementations of the research efforts. Advanced computing power and storage have allowed us to perform better data collection and simulations along with increased number of testing’s, the result of which can be seen in the sharp decline of error rates of mathematical and the thus derived computational models. The D.Ross model that we spoke about in the first para, now has five new variants, all of which supersede their prior models in terms of error rate, with the latest one being Wittikiend Model, which considers additional parameters such as Engine Power, No of Engines, Engine Mass etc., to compute the vessel noise. Moreover, the computational advancements allow us to perform novel analysis and superior demonstration of outputs, which in turn allows stronger data interpretation and implementation.

Artificial Intelligence – Machine Learning and its Impact on the UDA Framework

Artificial Intelligence (AI) is a very broad domain in itself and as such there is no single practical definition that covers all the aspects of AI. It is thus defined, depending on its usage in the respective field. To some it involves Automation and Robotics while to others it involves crunching Big Data and analyzing them. Nonetheless, the one thing that AI has common across all its facets is that theoretically, AI Refers to Machines, that are programmed to mimic human behavior such as learning and problem-solving. Furthermore, Machine Learning, which is a subset of AI, refers to this concept of Computer Programs, automatically learning from and adapting to new data and solving the given task, without being explicitly programmed by humans for each task. In practice, there are certain generic ML algorithms that are programmed to analyze any given set of suitable data and take rational decisions that have the best chance of achieving a specific goal, thus making the system to be Artificially Intelligent. This allows the system to be Scalable, Robust, handle Big Data and in some cases decrease the execution time of operation. Artificial Intelligence (AI) is therefore the most sought after technological advancement in the 21st Century with every industry adopting the same. But, what would implementation of AI for UDA look like?

"Any acoustic capacity building effort requires four processing stages: (a) Data Gathering, (b) Data Analysis, (c) Signal Processing and (d) Output, out of which the efficiency of second and third stage are heavily influenced by the quantity of data. This is where Machine Learning comes into picture."

Any acoustic capacity building effort requires four processing stages: (a) Data Gathering, (b) Data Analysis, (c) Signal Processing and (d) Output, out of which the efficiency of second and third stage are heavily influenced by the quantity of data. This is where Machine Learning comes into picture.

Mathematical Models used in Signal Processing algorithms are inherent to some implicit assumptions which decrease the accuracy unlike AI models which find patterns between data and output. Furthermore, the time complexity of mathematical models which basically involve matrix multiplication, linearly increases with increase in data points. AI models such as Neural Network which effectively perform Matrix Decomposition (Faster than Matrix Multiplication) require substantially less execution time. Lastly, any traditional mathematical models, hold true only for the environment in which they have been created. For e.g., the D. Ross model formed during Cold War falls short in prediction and is replaced by Wittikiend Model because there has been a substantial change in vessel dimensions over the years. Similarly, a few years down the line a new mathematical model will be required replacing the Wittikiend model and so on. AI models are robust in nature and with changes in data, it will automatically adapt to new situations.

"AI models are robust in nature and with changes in data, it will automatically adapt to new situations."

Ocean Ambient Noise is a crucial input parameter for signal processing in almost every underwater system and its estimation is therefore an integral part of any acoustic capacity building effort and the larger UDA. Contribution of AI towards UDA can be understood well using the following examples:-

Marine Spatial Planning is an application of UDA which aims at ocean ambient noise levels in the region, so as to provide information on areas which are hazardous to marine mammals. This is because, the frequency of communication of big marine mammals such as whales and dolphins is similar to the frequency in which shipping industry induces ambient noise and therefore the ambient noise creates hindrance to mammals, often resulting into their stranding and death.

Dr Christine Erbe and her team in 2011 mapped the cumulative shipping noise levels in the Vancouver region for Marine Spatial Planning. The mapping was performed by dividing the whole region into multiple grids and estimating shipping noise levels from static Automatic Identification System (AIS) data and estimating the impact of the ship’s noise level on each grid within a 100 Km radius, while accounting for suitable Transmission Loss. While this was an excellent effort using the then available models, it was a static map – without any real-time implementation possibility.
A more fruitful marine spatial planning using AI is performed by Google AI in August 2020, in order to protect Orcas (endangered killer whales) in the Salish Sea by reducing collisions. The AI model parses through the ocean’s soundtrack recorded using Hydrophones and then identifies Orca sounds to alert the Canadian Department of Fisheries and Oceans when the whales are present in the vicinity. This was accomplished by using a Convolution Neural Network (ML model) and training it on the processed signal image to distinctly segregate background noise from Orca Sounds.

Shipping Radiated Noise Estimation is another crucial aspect of acoustic capacity building since at low frequency 100-500 Hz, shipping is the most dominant source of anthropogenic noise in the ocean.

As it was mentioned earlier, D. Ross and Wittikiend are the two contrasting models for shipping radiated noise estimation. Even though Wittikiend model is proven to be more accurate than D. Ross for the existing vessels, it is the former one that is more popularly used because the input parameters of the Wittikiend model such as Number of Engines, Engine Mass and Engine Power etc., are difficult to obtain – more so in real time as they are not part of the AIS data set of a vessel.
An AI based approach easily solves the problem, as performed by MRC Pune, wherein a Random Forest Model was trained on the noise emitted from the ships recorded off the coast of Goa, which was then further matched with the Wittkiend Model output and programmed to predict the shipping noise levels, based on the length, speed and draught of the vessel. This approach retains the accuracy while reducing the number of required parameters as well as decreasing the overall execution time.

Cost of using Artificial Intelligence and what the future awaits

Big Data Better Results: ML algorithms needs data to learn and perform efficiently. This means, greater the size of data, better would be accuracy of the model. For building a robust model that could adapt to the changes in situations, one needs to be provided a continuous flow of data that would adequately represent these changes, and allow the model to ‘learn’. However, quantity of data is not the only factor for gaining accurate results, the quality of data being fed to the model is equally important. The data needs to be Complete, Unique, Valid and Consistent and this requires some pre-processing (data cleaning) before it can be actually used to train a model.

High Computational Power & Data Storage: Cleaning and analyzing the data to provide a quality input to the ML model is a computer intensive task. The bigger the size of data, greater the time required in data cleaning and therefore greater computational power and data storage. The bigger task, in addition to this is that of ML model operation itself. One needs to remember that the ML models operate on a trial & error method, searching for patterns within the input data points, and for this, it needs to iteratively analyze all data points with each other.

"From a bird’s eye view, implementing AI may seem to be a mammoth task for obtaining little increase in system efficiency, but what needs to be accounted for is not just the surge in efficiency but also robustness, allowing the model to adapt to ‘Change’, and ‘Change is the only constant in our ever progressing Maritime Domain. The future presents a tremendous scope for Artificial Intelligence in the UDA framework."

From a bird’s eye view, implementing AI may seem to be a mammoth task for obtaining little increase in system efficiency, but what needs to be accounted for is not just the surge in efficiency but also robustness, allowing the model to adapt to ‘Change’, and ‘Change is the only constant in our ever progressing Maritime Domain. The future presents a tremendous scope for Artificial Intelligence in the UDA framework. Some key projects that could revolutionize UDA are:

- Automated vessel movements with collision avoidance in the Inland Water Ways: The Inland Water Transport (IWT), networking throughout the country is a tremendous asset to India, which can substantially reduce transportation costs of goods and also mitigate carbon emissions to a large extent. Unfortunately, most of the IWT’s in India are currently not operational to its fullest, two of the reasons for which is stated as (a) Absence of proper Aids to Navigation and (b) Narrow Passages make it difficult for vessels to maneuver, especially if they are carrying large goods. Automating the vessel to be driver-less and mounting them with collision avoidance systems could be a pioneering work in solving the problem. Unlike humans, AI would mathematically plot thousands of movements at each instant and choose the best fit among them, thereby ensuring safe maneuverability.
- Dark Vessel Detection: Much of the illegal activities in the Open Ocean are performed by switching off their AIS, rendering them untraceable in AIS map. Over the years, there have been many technologies that make use of Synthetic Aperture Radar to detect such dark ships, but the limitation to which is that it cannot detect sub-surface vessels and is extremely costly to implement. A novel approach to solve the same using AI could use sensors placed in hot-spots, that would collect information of the noise level in the region and relay it to a ML algorithm, which will then automatically estimate the ‘ideal’ noise that should be present in the environment and check for anomaly between the estimated noise level and sensor recorded noise level. Presence of anomalies would suggest presence of another vessel in the vicinity that has not been recorded in the AIS. Running an image classification on the signal chart of recorded noise levels would then reveal the type of vessel. Deploying multiple sensors, would allow us to perform triangulation of the dark vessel, which may further provide its approximate location.
- Acoustic Habitat Degradation monitoring: The spatio-temporal low frequency ambient noise map generated using the AIS data as discussed above has multiple applications. The most prominent is the acoustic habitat degradation monitoring of big whales that perceive acoustic signal for biologically critical functions like foraging, navigation, communication, finding mates and more, in the same frequency band. Such inputs are extremely valuable for ensuring compliance of the Sustainable Development Goals (SDG-14) under the United Nations charter. Right from the policy at the strategic and tactical levels can be managed using this map.

Multiple such applications can be taken forward under the Digital Oceans initiative to ensure safe, secure, sustainable growth for all in the IOR. The MRC website gives more details on the same.