GAN-Based Sample Augmentation for Improving Above-Ground Biomass Inversion of Salt Marsh Vegetation Using Multi-Source Remote Sensing Data

Coastal wetlands are among the most efficient carbon sinks on Earth, acting as powerful natural engines that pull greenhouse gases from the atmosphere and lock them away in soil and plant tissue. These ecosystems, often called "blue carbon" habitats, include salt marshes where specialized plants thrive in salty, tidal conditions. To understand how much carbon these areas store, scientists must measure the "above-ground biomass," which is simply the total weight of all the living plant material—stems, leaves, and branches—growing above the soil. Traditionally, getting this number has been a slow, labor-intensive process requiring researchers to wade into muddy fields, cut down plants by hand, dry them in ovens, and weigh them. While accurate, this method can only cover tiny patches of land, making it difficult to get a complete picture of vast, remote wetlands.

To solve this problem, researchers are turning to satellites that can see the entire landscape at once, but these space-based tools need ground truth to learn how to read the data correctly. A new study from the Liaohe Wetland in China tackles the challenge of having too few ground measurements by using a clever artificial intelligence technique to create realistic "virtual" samples. By combining satellite images with a limited set of real field data, the team developed a method to estimate plant weight across the entire region with surprising precision. Their work offers a new way to monitor these critical ecosystems, helping scientists track carbon storage without needing to visit every single square meter of the marsh.

The Liaohe Wetland, located where the Liaohe River meets the Bohai Sea, is a vast expanse of tidal flats and salt-tolerant vegetation. It is home to two dominant plant types: the tall, reed-like Phragmites australis and the low-growing, salt-tolerant herb Suaeda glauca. These plants are vital for the region's carbon cycle, but their distribution is complex, shifting with the tides and the seasons. To map their biomass, the research team first gathered real-world data over several years, visiting specific spots in July, November, and September to measure the dry weight of the plants. They paired these measurements with images from the European Space Agency's Sentinel-2 satellites, which capture detailed views of the Earth's surface in various colors of light, including those invisible to the human eye.

The challenge was that the team only had 53 real measurement sites, a number too small to train a computer model to accurately predict biomass across the entire 128,000-hectare wetland. If a computer learns from too few examples, it often fails when faced with new, unseen areas. To overcome this, the researchers employed a Generative Adversarial Network, a type of artificial intelligence that works like a pair of competing programs. One program, the generator, tries to create fake data that looks exactly like the real measurements, while the other, the discriminator, tries to spot the fakes. As they play this game over and over, the generator gets better and better at producing synthetic samples that are statistically indistinguishable from the real ones.

Using this method, the team expanded their small dataset of 53 real samples into a much larger training set of 600 samples, with 300 for each plant type. These new, virtual samples were not random guesses; they preserved the exact statistical patterns and relationships found in the real world, such as how taller reeds generally correspond to specific signals in the satellite images. The researchers then trained a neural network, a brain-like computer model, on this expanded dataset to learn the complex, non-linear relationship between the satellite signals and the actual weight of the plants. They tested this approach against other common machine learning methods, finding that their neural network, fueled by the augmented data, consistently outperformed the others.

The results showed a significant improvement in accuracy. For the tall reeds, the model's ability to predict the correct weight improved substantially, with the error rate dropping by nearly 100 grams per square meter compared to models trained only on the original small dataset. For the smaller herbs, the improvement was also clear, with the model becoming much more reliable in estimating their biomass. When the team applied their best-performing model to the entire wetland using satellite images from September 2024, they produced a detailed map of plant weight across the landscape. The map revealed that the tall reeds, concentrated in the upper reaches of the estuary, held an average weight of 514.22 grams per square meter, while the smaller herbs in the lower, more open mudflats averaged 230.07 grams per square meter.

This study demonstrates that artificial intelligence can effectively bridge the gap between limited field work and the need for large-scale environmental monitoring. By generating realistic virtual data, scientists can build more robust models that work well even when ground measurements are scarce. The findings confirm that the Liaohe Wetland is a significant reservoir of carbon, with the tall reeds storing the vast majority of the biomass. This approach provides a reliable, scalable tool for assessing the health and carbon storage capacity of coastal wetlands, offering crucial support for global efforts to understand and protect these vital blue carbon ecosystems.