NOTE: The following blog is Part 2 of a five-part series discussing areas of the AI field that the Atlas AI team is passionate about and working to advance in 2024. If you missed Part 1 you can read it here.
Peering into the Future with Geospatial Foundation Models
Atlas AI was founded at the intersection of two major trends: the proliferation of geospatial data, particularly satellite imagery, and booming machine learning capabilities thanks to deep learning. Over the last few years, Atlas AI has demonstrated that this confluence of data and powerful analysis tools provides invaluable insights into Earth’s dynamics, climate patterns, and urban development.
The field of machine learning is now undergoing a new paradigm shift due to the emergence of foundation models (see our Stanford whitepaper on foundation models). Foundation models are a class of large-scale deep learning models that are pre-trained on vast datasets and can be adapted to a wide range of tasks. Their emergence has represented a new inflection point in the field of AI and machine learning because, unlike traditional models that are designed for specific tasks, foundation models can learn generalized representations of data, enabling them to perform well across multiple domains with minimal task-specific fine-tuning. This shift allows for more flexible, powerful, and efficient use of AI models in various applications.
Recent advancements in foundation models have predominantly centered around domains such as language, imagery, video, coding, and their multi-modal integrations, harnessing the power of large datasets to achieve remarkable versatility and efficiency in these fields. However, in our lab at Stanford and at Atlas AI, we are at the forefront of pioneering the development of geospatial foundation models. A geospatial foundation model is a type of large-scale deep learning model specifically trained on a wide array of geospatial data, including satellite imagery, topographical maps, and other location-specific datasets. This type of model learns to understand and interpret the complex patterns and relationships inherent in location data. Three particular areas of application that we’re working on are:
Forecasting Planetary & Societal Change
One particularly exciting area of application for these models is in the field of forecasting. By analyzing historical satellite data alongside other relevant information, geospatial foundation models can predict environmental changes, urban development, climate patterns, and even socio-economic trends with unprecedented accuracy and detail, offering invaluable insights for fields ranging from global logistics, infrastructure investment, and sustainable development. As an example, recent research from my group demonstrates that it is possible to leverage recent advances in diffusion models to forecast satellite-imagery time series. Going into 2024, I am particularly excited about Atlas AI playing once again a pioneering role in shaping foundation models as transformative tools to revolutionize the way we process, understand, and utilize geospatial data to peer into the future of our rapidly changing planet.
Enhancing Image Recognition
Another of the key roles of foundation models in geospatial data analysis is enhancing image recognition. Satellite imagery is often challenging for traditional algorithms (and human experts!) to interpret accurately. Inspired by recent advances in developing foundation models for natural images, my group at Stanford recently developed SatMAE, the first foundation model trained specifically on and for satellite imagery. Our research suggests that SatMAE excels in recognizing subtle patterns, identifying objects, and anomalies within these images, achieving new state-of-the-art results in a variety of benchmarks. In the coming year, I am excited to explore the application of these technologies within Atlas AI’s core analysis tools to better understand and monitor human and economic development across the world, ultimately learning from SatMAE’s initial applications to pave the way for Atlas to train the next generation of geospatial foundation model.
Contextualization of Geospatial Data
Another significant contribution of foundation models to geospatial data analysis is semantic contextualization. Leveraging their multi-modal capabilities, it will be possible to comprehend the context of satellite imagery, incorporating factors such as climate, social and economic data which could be either in tabular or text formats. By recognizing patterns across massive amounts of imagery, text and tabular data, these models can understand the context inherent in a particular imagery, providing faster and more insightful alerts about (a) what is happening within the image, (b) what has changed about the image, and (c) what is relevant about the image for a particular operational objective.
In this way, incorporating foundation models into geospatial data analysis should streamline decision-making processes. Models will be queried directly in natural language and directly generate detailed reports and summaries based on the analyzed data, enabling policymakers, scientists, and researchers to make well-informed decisions.
Conclusion
Foundation models are ushering in a new era in geospatial data analysis. Their ability to enhance image recognition, enable predictive analysis, provide semantic contextualization, and facilitate decision-making processes is going to transform the field and we’re proud that Atlas AI continues to be at the center of this revolution.
Stefano Ermon is a Co-Founder of Atlas AI and an Associate Professor in the Department of Computer Science at Stanford University, affiliated with the Artificial Intelligence Lab and the Woods Institute for the Environment. His research focuses on machine learning and generative AI, with a keen interest in developing principled methods driven by real-world applications and broad societal issues.
Professor Ermon's teaching includes courses like Probabilistic Graphical Models and Deep Generative Models. His work has earned him several honors, including the ICLR Outstanding Paper Award, the Sloan Research Fellowship, and the NSF CAREER Award, reflecting his impactful contributions to AI and machine learning.
