Key finding
The team built a convolutional neural network (VGG16-U-Net) trained on more than 23,000 Planet Dove images to create annual 3-meter resolution kelp canopy maps for the entire California coast from 2017 to 2024. The model achieved 89.6% overall accuracy and was validated against UAV surveys (R2 = 0.86) and aircraft surveys from the California Department of Fish and Wildlife (R2 = 0.80). The high-resolution maps revealed patterns invisible at 30 meters.
Kelp persistence was most strongly predicted by historical persistence before the heatwave. Cooler sea surface temperatures, shallower depths, and lower habitat fragmentation also predicted higher persistence. But these effects varied dramatically by latitude: what mattered in Southern California was different from what mattered in the north. The paper released all mapping data publicly through the KelpWatch platform (kelpwatch.org).
What the research asks
The marine heatwave of 2014-2016 devastated California's bull kelp forests, causing a collapse in canopy cover of over 90% in some regions. But even in the midst of this broad-scale loss, some kelp patches persisted while others disappeared. Understanding why some forests survived while others did not is essential for restoration planning, but most satellite monitoring has been limited to 30-meter resolution, too coarse to detect the fine-scale local factors that determine survival. This study asked whether deep learning on 3-meter resolution Planet Dove imagery could identify the local drivers of kelp canopy persistence after a major marine heatwave, producing the first annual high-resolution kelp canopy maps for the entire California coast.
What it finds
The team built a convolutional neural network (VGG16-U-Net) trained on more than 23,000 Planet Dove images to create annual 3 m resolution kelp canopy maps for the entire California coast from 2017 to 2024. The model achieved 89.6% overall accuracy and was validated against UAV surveys (R2 = 0.86) and aircraft surveys from the California Department of Fish and Wildlife (R2 = 0.80).
The high-resolution maps revealed striking patterns invisible at 30 m. Kelp persistence was most strongly predicted by historical persistence before the heatwave, meaning areas that had long been stable were most likely to recover. Cooler sea surface temperatures, shallower depths, and lower habitat fragmentation also predicted higher persistence. But these effects varied dramatically by latitude: what mattered in Southern California was different from what mattered in the north. This regional variability underscores the need for local-scale data in marine management, not one-size-fits-all approaches.
The paper also released all mapping data publicly through the KelpWatch platform (kelpwatch.org), making it accessible to restoration practitioners, fisheries managers, and the public.
Why it matters
This paper showcases Planet imagery applied to blue carbon and marine biodiversity monitoring, diversifying beyond the forest-focused applications that dominate the earth observation sector. The use of deep learning to extract actionable ecological information from high-resolution imagery is a model that can be extended to other coastal ecosystems: seagrasses, mangroves, and coral reefs.
The Planet connection is explicit throughout: the paper is titled after Planet Dove data, relies entirely on Planet's CubeSat constellation, and demonstrates that commercial high-resolution imagery can answer questions that free medium-resolution data cannot. The 3-meter resolution was essential for detecting the fine-scale patterns of kelp persistence that 30-meter Landsat data simply could not resolve. For marine conservation practitioners and blue carbon project developers, this study provides an open-access methodology and data product for monitoring kelp forest recovery, supporting both biodiversity reporting and carbon accounting.