Mapping Kelp Forests
Rank: 1st out of 671
Helping researchers track the health of Giant Kelp Forests by segmenting Satellite Imagery
About the competition
In this competition hosted by the Woods Hole Oceanographic Institution (WHOI), Team Epoch will use machine learning analysis of satellite imagery to map the presence and absence of giant kelp forests. These crucial canopies, visible even from space, will be classified through an AI model, enhancing the scope and cost-effectiveness of the research.
Woods Hole Oceanographic Institution
The WHOI is an independent non-profit organization dedicated to improving ocean research, exploration, and education. Together with research institute the Zooniverse and scientists from UMass Boston, WHOI is hoping to find a model that can classify giant kelp with accuracy comparable to the human eye.
Relevance
Coastal ecosystems face mounting threats from factors such as climate change and overfishing. Giant kelp plays a crucial role in the sea’s ecosystem. It is recognized as a “foundation species” as it provides food for various herbivores, enhances ocean health, and acts as a highly efficient carbon sink. Its global distribution is widespread, and it is one of the most easily identifiable forms of kelp in satellite images.
Mapping life on the seafloor is crucial for scientists. This is generally done by scuba divers getting samples from the deep. Doing this is both costly and time-intensive.
A more efficient approach for mapping and monitoring
critical ecosystems is using a machine learning analysis of
satellite imagery.
Technical details
The dataset used in this competition relies on imagery from Landsat satellites, which photograph the entire Earth’s surface every 16 days. The challenge of the competition is to create an algorithm that predicts the presence or absence of kelp canopy using these satellite images. The satellite data consists of multiple bands, such as Shortwave-Infrared, Near-Infrared and RGB. Next to that we are provided with a cloud mask and an elevation map of the terrain. Models that are commonly used in this computer vision task consists of large pre-trained convolutional neural networks (CNN) which leverage the knowledge and feature representations learned from diverse satellite images. These can then be fine-tuned on the kelp satellite data using transfer learning. Vision Transformers have also made their entrance and are competitive or even superior compared to CNN-based architectures due to their ability to capture global context information thanks to their
self-attention mechanisms.
UN Sustainable Development Goals
This competition is in line with the UN Sustainable Development Goals 13: Climate Action and SDG 14: Life Below Water. Improvement of research on giant kelp forests will increase localized marine biodiversity and ecosystem resilience. Furthermore, it improves the capture of CO2.
