Clearing the Air with AI

When sunlight passes through the atmosphere and bounces off the Earth, it is affected by factors like turbulence, temperature changes, and various gasses, which distort the data collected by remote sensing satellites. This impacts a wide range of important applications, from generating drought and vegetation indices for monitoring photosynthetic activity, to the detection of methane plumes and even activity at foreign military bases. Scientists have developed methods to correct this distortion, known as atmospheric correction, by understanding how the atmosphere affects sunlight.

However, traditional atmospheric correction methods use generic profiles and are either time-sensitive or high-cost and data-intensive. But a new approach developed at the Pacific Northwest National Laboratory strikes a balance between these extremes, providing good accuracy with less data. This was achieved by using physics-informed machine learning to enhance remote sensing, which makes it possible to automatically correct atmospheric distortions and predict how materials on the ground would look through different atmospheric conditions.

Under normal circumstances, dealing with all of the possible sources of variation would require a staggering amount of training data to produce an accurate deep learning model. But physics-informed machine learning sidesteps this requirement by allowing researchers to use prior knowledge to help with the training of the neural network. This greatly enhances the efficiency of the process and also makes it more accurate. In this case, the team was able to train their model using little more than a hundred satellite images.

Instead of starting with a blank slate, as is the case with most neural networks, this algorithm has a set of differential equations at its core. This constrains it, such that it only seeks potential solutions that conform to what we already know about how sunlight is altered as it passes through the atmosphere, and after it bounces off of a target like the ground. This type of solution might not make sense where we have little knowledge of the phenomena we are trying to model, but in this case, we do not need it to understand the physics for us — we just need it to quickly perform calculations that are otherwise very complex.

To demonstrate the effectiveness of this approach, the researchers leveraged their pipeline to learn the profile of the atmosphere and coastal waters. This created a new avenue for monitoring the health of coral reefs from satellites. At present, the team is further refining their methods with the hopes of improving the algorithm’s performance. They are also exploring other potential applications of their system where existing data sources are very limited.