Kelp Wanted: Segmenting Kelp Forests

Help researchers estimate the extent of Giant Kelp Forests by segmenting Landsat imagery. #climate

$15,000 in prizes
Completed feb 2024
671 joined

Overhead drone footage of giant kelp canopy. Image Credit: Tom Bell, All Rights Reserved.


“Underwater kelp forests nurture vibrant and diverse ecosystems around the world. Kelp forest health can be threatened by marine heat waves, shifts in grazers, poor water quality (often due to run off linked to changes in nearby landscapes), and over-harvesting. These pressures can manifest locally or globally, and with complicated spatial dynamics. For example, we've recently witnessed forests struggling in some regions, while thriving in others, and we still don't fully understand why.

Developing new tools to use satellites for monitoring and understanding kelp forests is critical for us to understand where and how kelp forests are changing.”

— Dr. Henry Houskeeper, Postdoctoral Scholar at Woods Hole Oceanographic Institution and a primary contributor to KelpWatch.org

Why

Kelp forests are crucial underwater habitats that cover vast swaths of ocean coastlines. Giant kelp, the linchpin of these ecosystems, provide shelter and food for many species and generate substantial economic benefits. Despite their importance, kelp forests face mounting threats from climate change, overfishing, and unsustainable practices.

Better tools are needed to monitor and preserve kelp forests on a large scale. However, monitoring these habitats can be difficult because they change rapidly in response to temperature and wave disturbances. Satellite imagery provides an opportunity to monitor kelp presence often, affordably, and at scale.

The Solution

In the Kelp Wanted challenge, participants were tasked with developing semantic segmentation algorithms to predict the presence or absence of kelp canopy in satellite imagery of coastal waters surrounding the Falkland Islands. The challenge provided 350 x 350 pixel "tiles" of Landsat satellite imagery and labels generated by citizen scientists from the Floating Forests project. Winning algorithms will not only advance scientific understanding, but also equip kelp forest managers and policymakers with vital tools to safeguard these vulnerable and vital ecosystems.

The Results

Over the course of the competition, participants battled it out and submitted a whopping 2,885 solutions! The top of the leaderboard remained fiercely contested throughout the competition, and the winners were ultimately able to achieve impressive results. The first place team secured the top spot with a Dice Coefficient of 0.7332. Close on their heels, second and third place scored 0.7318 and 0.7296, respectively. These results signify exciting advancements in the ability to accurately detect the presence and extent of kelp on a large scale.

To see the winners' models and write-ups, see the publicly accessible winners' repository for this competition.


RESULTS ANNOUNCEMENT + MEET THE WINNERS

WINNING MODELS ON GITHUB