Built in collaboration with IBM and NASA, the watsonx.ai model transforms satellite data into high-resolution maps of floods, fires, and other topographical changes designed to reveal Earth’s past and suggest its future. It has been.
Nearly a quarter of the world’s population now lives in flood zones, and that number is expected to rise as climate change raises sea levels, intensifies storms, and puts more people at risk. I’m here. The ability to accurately map flood events not only protects people and property now, but could be key to directing development to lower-risk areas in the future.
A new geospatial foundation model announced today by IBM is designed to enable a first step toward this goal by transforming NASA satellite observations into customized maps of natural disasters and other environmental changes. It has been. Part of IBM’s watsonx.ai geospatial offering, the model will be available in preview to IBM clients through the (EIS) IBM Environmental Intelligence Suite in the second half of this year. Potential uses include estimating climate-related risks to crops, buildings, and other infrastructure; assessing and monitoring forests for carbon offset programs; and creating strategies for companies to mitigate and adapt to climate change. development of predictive models to help
As part of the Space Act Agreement with NASA, just four months ago, IBM set out to build the first ever foundational model for analyzing geospatial data. The Foundation Model has revolutionized natural language processing (NLP). A developer can train one model on his raw text, and with additional training customize the model for his other NLP tasks. Previously, users had to train a new model for each task, which required extensive data curation and computation. Instead of training the underlying model in words, IBM Research taught the model to understand satellite imagery. Researchers pre-trained on NASA’s Harmonized Landsat Sentinel-2 (HLS-2) data. HLS data are consistent from the Operational Land Imager (OLI) aboard the joint NASA/USGS Landsat 8 satellite and the Multi-Spectral Instrument (MSI) aboard the European Union’s Copernicus Sentinel-2A and Sentinel-2B satellites. Provides surface reflectance data. The combined measurements enable global observations of land with a spatial resolution of 30 meters every 2-3 days.
We then fed hand-labeled examples into the model and trained it to recognize the extent of historical flood and fire burns, changes in land use and forest biomass, and more.
The model is designed to be easy to use, as users only need to select a region, a mapping task, and a set of dates. For example, if a user types “Port-de-Lanne, France” into the search bar and selects dates from December 13-15, 2019, the model highlights in pink how far the flood has spread. To do. Users can overlay other datasets to see where crops and buildings have been flooded. These visualizations are useful for future planning in similar disaster scenarios. They provide information to help mitigate the effects of flooding, inform insurance and risk management decisions, plan infrastructure, respond to disasters, and protect the environment.
IBM built a model with masked autoencoders to process video and adapted it to satellite imagery. To make the model understand a sequence of images that unfolded over time, the researchers erased a portion of each image and had the model reconstruct it. The more images we reconstructed, the better we were able to understand how they relate to each other. We then fine-tuned the model for specific tasks such as image classification and segmentation. This fine-tuning workflow was based on her PyTorch with an enhanced segmentation library that allows researchers to work with spatiotemporal data.
To improve the model’s efficiency, the researchers also reduced the size of the satellite imagery. This allowed us to process data in smaller chunks and use fewer GPUs. He then used over 5,000 GPU hours to train the model on IBM Research’s Vela supercomputer.
Early results look promising. In tests, researchers found a 15% improvement in accuracy using half-labeled data compared to state-of-the-art deep learning models for mapping flood and fire scars. IBM estimates that the model can speed up geospatial analysis by 3-4x and reduce the amount of data cleaning and labeling required to train traditional deep learning models.
See how this technology can be applied for businesses looking for easier and faster ways to analyze climate data and derive insights. For example, disaster response teams can use such solutions to prepare for fires that affect homes. Alternatively, the solution helps large consumer goods companies better understand macro trends such as climate change, severe weather and geopolitical risks. These trends will impact where we are buying raw materials today and where we might consider buying those resources in the future. It also helps reduce the impact of agricultural practices on the local environment and surrounding communities by better understanding how large-scale agribusiness can mitigate pollution caused by soil degradation, water conservation activities, or field runoff into the area. It also helps you better measure, track, and mitigate. body of water.
IBM is also involved in additional projects on geospatial mapping, working with Esri and others to help organizations uncover a richer spatial context for their business assets and operations.
Try the latest in geospatial AI
A preview version of the base geospatial foundation model and a set of fine-tuned models, all running on watsonx.ai, will be available through IBM EIS.1 We believe this release will be invaluable for data scientists, developers, researchers, and students.
A preview trial will make these two sets of tools available within EIS. It includes APIs for inferring fine-tuned models that drive developer engagement. We also provide sample solutions that apply these models, create triggers, and leverage work queues to drive downstream operational systems.
If you are interested in joining the waitlist and testing the preview version within IBM EIS, click here.
