AI models trained on NASA satellite data aid disaster response

When Hurricane Ida hit Louisiana as a category 4 storm on August 21, 2021, NASA's Disasters Program supplied responders with remotely sensed data rather than boots on the ground. The program, part of NASA's broader commitment to open science, produced maps and models tracking soil moisture changes, vegetation changes, rainfall totals, flood extent, and nighttime lights, the last of which helped identify power outages. That data was shared through the NASA Disasters Mapping Portal and used in cross-agency coordination calls.

One specific application was detecting blue tarps on rooftops in Ida's aftermath. Blue tarps are commonly used to cover storm-damaged roofs, so counting them from satellite images gives a rough measure of how badly a community was hit. The Interagency Implementation and Advanced Concepts Team, a group based at NASA's Marshall Space Flight Center in Huntsville, Alabama, used artificial intelligence on satellite images to conduct this count.

NASA also overlaid publicly uploaded photos onto its Damage Proxy Maps, giving responders a view of ground conditions before, during, and after the storm.

Beyond single-disaster tools, NASA and IBM are jointly developing five open-source AI foundation models trained on NASA's satellite archives, an effort led by NASA's Office of the Chief Science Data Officer. A foundation model is trained once on a broad pool of data and can then be adapted to many different tasks, rather than being built from scratch for each one. NASA describes potential uses of these models, called the Prithvi Earth Foundation Models, as ranging from identifying flood risk and predicting crop yields to forecasting long-range atmospheric weather patterns.

Kevin Murphy, NASA's chief science data officer, said the models are designed to be scientifically validated and adaptable to new data, with the goal of lowering technical barriers so response teams can act quickly even during difficult disasters. NASA frames the effort as part of a wider push to make its growing volume of satellite data more usable to outside groups.

Outside NASA, research groups have applied a range of machine learning methods to forecasting and detecting disasters. A review published in Frontiers in Environmental Science found a steady rise since 2014 in studies using algorithms such as neural networks, decision trees, and random forests to predict weather patterns, and a similar rise in studies applying machine learning to natural disaster prediction generally, with 2022 showing the most such research in the period the review covered. These models draw on satellite imagery, atmospheric data, and historical weather and disaster records to find patterns humans might miss, according to the review, which cites NOAA's Global Ensemble Forecast System as an example of an MLA-based model that draws on atmospheric data and historical weather data.

Separate reporting from IEEE Public Safety Technology describes similar techniques feeding early warning systems for hurricanes, earthquakes, floods, and wildfires, using real-time data from weather stations, satellite imagery, and seismic sensors. It describes convolutional neural networks being used for hurricane trajectory prediction, recurrent neural networks for identifying tremors that may precede earthquakes, and support vector machines for classifying flood and wildfire risk from factors like rainfall, river levels, and vegetation density. A Nature collection on the subject notes that these disasters affect roughly 160 million people worldwide every year and that machine learning's growing traction stems partly from shrinking processing timelines, growth in computational power, and increasing usability of large data sets.

A separate study, published in PubMed Central and not yet noted as peer reviewed in the source text, trained deep learning models on a newly compiled set of 6,334 aerial and satellite images to detect flooding and desertification, a category the researchers say had rarely been included in prior disaster-detection datasets. Of four models tested — a convolutional neural network, DenseNet201, VGG16, and ResNet50 — DenseNet201 and ResNet50 reached the highest testing accuracies, at 99.37% and 99.21% respectively, on that dataset. The researchers describe drone-based emergency response as a practical application of the work.

None of the sources describe these systems as infallible. The IEEE overview states plainly that despite improved accuracy, machine learning models