Mapping Coastal Forest Retreat Using Convolutional Neural Networks and Different Satellite Imagery
This study utilizes convolutional neural networks and multi-source satellite imagery to demonstrate that incorporating phenological and topographical indices improves ghost forest detection, revealing that 21% of North Carolina's coastal forests were lost between 1985 and 2021 due to accelerating sea-level rise, salinity, and proximity to channels, with the rate of conversion to marshes and ghost forests increasing significantly in the last decade.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Along the low-lying coast of the eastern United States, a quiet transformation is reshaping the landscape. For centuries, saltwater has met freshwater forests at the edge of the land, but rising seas and shifting tides are pushing that boundary inland. When saltwater invades soil that trees are not built to handle, the trees do not simply fall over; they stand still, their leaves turning brown and their branches going bare, creating a silent, skeletal landscape known as a "ghost forest." Over time, these dead trees are replaced by salt-tolerant grasses and shrubs, and eventually, the land becomes open marsh or water. This process, driven by the slow creep of the ocean and the increasing salinity of the ground, is a visible sign of climate change, but tracking exactly where and how fast it happens has been difficult. Scientists need to know not just that the forests are disappearing, but which specific factors—like how close a tree is to a water channel or how high the ground sits—are causing the death of the trees, so that conservation efforts can be directed where they are needed most.
To solve this puzzle, researchers turned to a combination of satellite imagery and advanced computer learning. They focused on a vast stretch of coastal North Carolina, an area known for its rapid rise in sea level and its mix of freshwater swamps, upland forests, and salt marshes. The team used two different types of satellites to watch this landscape over time. One satellite, Landsat, has been circling the Earth for decades, providing a long historical record of the land, though its images are somewhat blurry. The other, Sentinel-2, is newer and captures much sharper, more detailed pictures, but it has only been around for a shorter time. The researchers wanted to see if they could use these images to teach a computer how to spot the subtle differences between healthy green forests, the dead gray of ghost forests, and the shrubby vegetation that often grows in between. They knew that looking at the land at different times of the year was crucial, because healthy trees follow a predictable cycle of greening up in spring and losing leaves in winter, while dead trees and shrubs do not follow the same rhythm.
The scientists built a computer model that could learn to recognize these patterns. Instead of just looking at a single snapshot of the land, they fed the model data that tracked how the vegetation changed throughout the seasons. They also included information about the shape of the land, such as its height and slope, and how close it was to rivers and canals. By training the computer with thousands of examples of what healthy forests, ghost forests, and marshes look like, they created a system that could map these changes with remarkable precision. When they tested the model, the sharper Sentinel-2 images performed slightly better than the older Landsat images, correctly identifying the land cover types in nearly all cases. However, the researchers found that the long-term data from Landsat was still essential for understanding the full story of how the landscape has changed over the last few decades.
Using the best data available, the team mapped the coastal plain from 1985 to 2021. The results revealed a dramatic shift. Over these thirty-six years, the region lost about 21 percent of its forest cover. In the earlier part of the study, from 1985 to 2010, the forest loss was significant, but the pace accelerated sharply in the last decade. Between 2010 and 2021 alone, nearly 24,000 hectares of forest were converted into marsh, ghost forest, or shrubland. This rate of loss was one and a half times higher than the loss seen in the previous twenty-five years. The study showed that this transition is not happening randomly; it is driven by specific environmental pressures. The closer a forest is to a water channel, the more likely it is to die, as these channels act as pathways for saltwater to move inland. High levels of salt in the soil and the accelerating rate of sea-level rise were the strongest predictors of where ghost forests would appear.
The researchers also discovered that the landscape is changing in complex ways that go beyond simple forest-to-marsh conversion. While many trees are dying and turning into ghost forests, others are being replaced by shrubs, and some areas are even seeing new forests grow on land that was once farmland. However, the overall trend is one of retreat. The study confirmed that the loss of forest is concentrated in low-lying areas near the coast, but it is also pushing further inland along drainage ditches and canals. The findings suggest that as sea levels continue to rise and extreme weather events become more frequent, the pressure on these coastal ecosystems will only intensify. By providing a clear, detailed map of where these changes are happening and why, the study offers a vital tool for planners and conservationists. It highlights the specific regions most vulnerable to the encroaching salt, allowing for targeted efforts to protect what remains of these forests or to manage the transition to new, salt-tolerant ecosystems before it is too late.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.