A new artificial intelligence (AI) model called Aquila helps computers better understand remote sensing images of the Earth, taken from satellites and airplanes. Aquila combines clear, high-detail image processing with improved language abilities so it can describe scenes more accurately and answer complex questions about what it sees.
Earlier systems often struggled with overhead images because roads, buildings, and fields can look very different at various heights and resolutions. Many models lost important small details when connecting pictures to words. Aquila addresses these issues by accepting much larger images, up to 1,024 by 1,024 pixels, and by mixing visual information from several scales repeatedly during its reasoning process.
Improved design for remote sensing challenges
The model uses three main parts: a special vision encoder based on ConvNeXt, a module that brings together features from different scales while keeping spatial structure, and a language model based on Llama-3 with deep alignment between images and text. This approach preserves fine details that are crucial in remote sensing, such as the layout of fields or the shape of coastlines.
In tests, Aquila performed noticeably better than previous models. On one benchmark for describing images, it improved results by nearly 8 percent. On a visual question answering task, it reached almost 84 percent accuracy. It also showed strong ability to point out specific objects in images. The system was trained in two stages using millions of remote sensing image-text pairs on standard graphics processing units.
Aquila remains computationally demanding and currently works only with single-time RGB color images. However, its design provides a foundation for future systems that could handle multiple time periods, different types of sensor data, or radar images. The advance makes Earth observation data more useful for tasks such as tracking crops, monitoring cities, managing disasters, analyzing the environment, and supporting better planning decisions. By keeping spatial precision while using natural language, Aquila brings expert-level understanding closer to everyday users of satellite imagery.
This research is published in Journal of Remote Sensing.