Novel deep learning architectures and problem formulations for the geosciences.
While data science and deep learning have achieved tremendous progress for image classification and segmentation recently, geoscience problems have not reaped the benefits yet. I discuss some fundamental differences between common machine learning tasks and datasets and typical geoscience data, in terms of scale, the number of examples, and ground truth/label availability. As part of this discussion, I highlight opportunities related to the non-real-time nature of various geoscience problems and show how data science techniques can benefit.
Next, I discuss what properties of geoscientific data justify my focus on deep learning, and show that deep learning is particularly useful when there are no physical models for a certain earth scientific task. Another key target problem is answering specific questions about complicated datasets that conventionally rely on generating large intermediate earth models. A guiding example throughout this talk is the mapping of aquifer types from surface observations and airborne/satellite geophysical and remote sensing data.
One of the main limiting factors when training deep neural networks on geoscience data is the memory required for network parameters and network states. Both grow with network depth. I present new network designs and training methods that change linear memory growth into a constant. The new methods are based on the physics of wave propagation and related geophysical parameter estimation procedures. The saved memory enables larger data inputs, so our networks can learn from larger structures in the data.
Besides aquifer mapping, I show examples from land-use change detection using hyperspectral data, surface water chemistry prediction, geological model building from borehole & seismic data, as well as geological mapping & mineral prospectivity from airborne geophysical observations.