Oceans are the driving force of Mother Nature, holding 97% of earth’s water. Oceanic ecosystems involve many critical marine species such as fishes, seagrasses, and coral reefs. These are essential in the marine ecosystem, for example, if seagrasses are removed, this may lead to the reduction of light required for photosynthesis.
At the same time, it involves huge maintenance of these marine species. Due to tourism, shipping, and human intervention, 75% of the world’s coral reefs are being threatened and 19% of the coral reefs having been destroyed by 2011.
Managing these destructive impacts in a sea is an arduous task. Technologies were developed and are being developed to monitor the marine ecosystem. In recent years, the use of digital cameras, Unmanned Underwater Vehicles (UUV), and Autonomous Underwater Vehicles (AUV) has led to an exponential increase in the availability of underwater imagery.
reason to use deep learning technique
One of the technologies, Integrated Marine Observing System (IMOS) observes and collects millions of images of coral reefs around Australia, but less than 5% go through the expert marine analysis. For this reason, it is now a research priority to analyze marine digital data automatically.
Many Machine Learning techniques have been imposed to do analysis automatically. For the specific task and for specific data, with careful extraction of hand-crafted features, Machine Learning achieves good performance. Techniques such as Support Vector Machine (SVM), Linear Discriminator Analysis (LDA), and other conventional machine learning techniques get saturated quickly when there is a huge volume of training data.
In general, Machine Learning cannot handle a large amount of data and hand-crafted features have to be given to Machine Learning algorithms which become tedious when there are more features in the dataset. To solve these issues, Geoffrey Hinton developed Deep Neural Networks. (Read more about Geoffrey Hinton)
Deep Neural Networks (DNN) was developed and inspired by neurons in the human brain. Deep Learning has been proven to be state-of-art in many fields. Unlike Machine Learning, Deep Learning learns features automatically and there are no hand-crafted features are required which reduces a lot of manual work.
Deep Learning can be used for both supervised (with the label – a dog or not a dog) and unsupervised learning (without label and learning hidden patterns and structure on its own).
Some of the real-time application of DNN includes self-driving cars, speech recognition, image recognition, analysis and future prediction, and many more.
deep learning in marine Ecosystem
Deep Learning is one of the hottest technique used in many AI industries. But AI is still at it’s infant stage in marine ecosystem.
Many kinds of research are being continued to do automatic analysis and to maintain the marine ecosystem using deep learning and AI. Below are some of the use-cases where deep learning can help in a marine ecosystem:
Deep Learning for analysis and classification of coral reefs:
Healthy coral reefs play a vital role in the marine ecosystem. Millions of images have been collected with AUVs but require manual analysis which is not only time consuming but each image has to be observed and processed manually by a marine expert.
Automated technology allows for transformational ecological outcomes which helps in monitoring the health of the oceans.
CHALLENGE: There are many factors that are involved in coral reefs. The size, texture, color may vary differently. Moreover, things like algae blooms can change the turbidity of the water , affecting the color of the image. This kind of challenge makes conventional algorithms inappropriate.
This is where the Deep Learning comes into picture which can work better than conventional algorithms.
As mentioned earlier, a huge amount of images collected from AUVs has to be annotated and classified manually by marine scientists. So a technique called Convolution Neural Network (CNN) is used which can automatically annotate and classify coral reefs.
CNN is a deep learning algorithm that can learn hidden features automatically. CNN has three layers convolutional layer, pooling layer, and fully connected layer. (There are many mathematical concepts in CNN, but those are beyond this scope)
To annotate images, a feature extraction scheme based on combination of Spatial Pyramid Pooling (SPP) and CNN is used.
For texture and color characterization, Local Binary Pattern (LBP) and Normalized Chromaticity Coordinate (NCC) are used along with a three-layer Neural network for coral reef classification.
For high-resolution images, Depixelate Super Resolution CNN (DSRCNN) can be applied. The low-resolution image is reconstructed to form a high-resolution image that can then passed to classifier for classification.
Through these methods, analysis and monitoring of coral reefs underwater can be done easily. This would be of great benefit to know about decreasing trends in the population of coral reefs.
Deep Learning in Plankton classification:
Planktons are the base for food webs in the marine ecosystem. The same CNN architecture technique has been designed and used for the plankton classification. But an advanced technique called “roll operation with cyclic pooling” is applied.
The architecture involves 16 layers in total. Cyclic pooling provides the network the ability to use the same feature extraction from the input at different orientations, where the same stack of convolutional layers are applied and fed to dense layers and the feature maps are pooled together.
The stacks of convolutional layers are combined to form one large stack and linked to two fully connected layers for classification. But when the dataset is not large enough, this method was seemed to prone to overfitting and so small kernels (functions) are modified in the last two fully connected layers for such dataset.
The CNN with cyclic pooling achieved better performance in identifying and classifying plankton species. Even microscopic species could be identified with Deep Learning technique.
Like these, deep learning can be used for other recognition tasks such as live fish recognition and for many tasks in marine ecosystem.
- Deep Learning requires a large amount of dataset to be trained and validated.
- To get an underwater object dataset, high level integrated AUVs are required which should have a longer life span.
- To train Deep Learning, high-level computing hardware with Graphical Processing Units (GPU) like NVIDIA GPU’s are required.
Maintaining and monitoring the Marine Ecosystem is a tedious task. Deep Learning would greatly benefit in terms of cost, speed, and accuracy and in better quantifying the environmental change. By using AI techniques as Deep Learning can lead to a potential increase in the research process of ocean observation.
What are the advanced deep learning techniques do you think it can be used in marine ecosystem ?