Mapping of Land Use and Land Cover (LULC) using EuroSAT and Transfer Learning
This paper demonstrates that applying transfer learning and fine-tuning to RGB bands from the EuroSAT dataset enables highly accurate (99.19%) land use and land cover mapping, providing a valuable tool for environmental monitoring, urban planning, and climate change mitigation.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Earth is constantly changing, reshaped by the slow creep of forests, the sudden spread of cities, and the shifting patterns of agriculture. To manage these changes and protect our natural resources, scientists need a clear, accurate picture of what the ground looks like from above. For decades, this meant sending people into the field or manually studying photographs, a slow and often incomplete process. Today, satellites orbit the planet, capturing vast amounts of digital images every day. These images act as a massive, global archive of our surface, but they are too numerous for humans to sort through one by one. This is where artificial intelligence steps in, specifically a branch known as computer vision, which teaches machines to recognize patterns and identify objects within images just as a human eye would. The challenge lies in training these machines to distinguish between a river, a forest, or a residential neighborhood with high precision, especially when the data comes from complex satellite sensors.
In a recent study, researchers Suman Kunwar and Jannatul Ferdush tackled this challenge by testing how well a modern type of artificial intelligence could classify land use and land cover. They focused on a specific collection of satellite images called EuroSAT, which contains 27,000 labeled pictures of the Earth's surface taken by the Sentinel-2 satellite. These images are divided into ten distinct categories, ranging from annual crops and forests to highways and industrial zones. The researchers wanted to see if they could use a powerful, pre-trained model known as a Vision Transformer, or ViT, to learn how to sort these images. This model had already been taught to recognize objects in general photographs, and the team aimed to see if it could be adapted to understand the specific textures and colors of satellite imagery. They tested whether feeding the model extra, slightly altered versions of the same images—such as flipped or cropped copies—would help it learn better, a technique known as data augmentation.
The team trained their system using a specific set of instructions and tools, running the experiments on powerful computer processors available through an online platform. They compared the performance of the Vision Transformer against two other well-known artificial intelligence models, VGG16 and ResNet-50, using the same satellite data. The results showed that the Vision Transformer was the most accurate of the three. When the researchers used the standard images without any extra alterations, the model correctly identified the land type in 98.61 percent of the test cases. However, when they introduced the data augmentation technique, the model's performance improved further, reaching an accuracy of 98.72 percent. In a more detailed test where they trained the model for a longer period, the version using augmented data achieved a remarkable 99.19 percent accuracy, while the version without augmentation saw its accuracy decline slightly as training continued, dropping from 98.67% at 10 epochs to 98.41% at 20 epochs. This suggests that exposing the model to a wider variety of image angles and crops helped it generalize better to new, unseen data.
To prove that this high level of accuracy could be applied to the real world, the researchers took their best-performing model and applied it to a specific region known as Kreis Borken in Germany. They gathered satellite images of this area from 2018 to 2020 and broke them down into small, manageable squares. The model then examined each square and assigned it a label, such as forest, pasture, or residential area. The final output was a colorful map that visually represented the distribution of different land types across the region. The analysis of the map revealed that the model was particularly good at identifying forests and bodies of water, while it found the "pasture" category slightly more difficult to distinguish, recording the lowest accuracy among all classes, though the remaining categories achieved nearly 99% accuracy.
The study concludes that using transfer learning, where a model trained on one task is adapted for another, is a highly effective way to map land use. The researchers found that adding data augmentation not only improved the accuracy of the classification but also helped the model avoid overfitting, a common problem where a computer memorizes the training data too closely and fails to recognize new patterns. By combining these techniques, the team demonstrated that it is possible to create detailed, automated maps of the Earth's surface with near-perfect precision. These findings suggest that such tools can be valuable for urban planners and environmentalists who need to monitor changes in land use, plan for future growth, and protect natural habitats without relying solely on manual observation.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.