Coupled Statistics-Physics Guided Learning to Harness Heterogeneous Earth Data at Large Scales

Despite recent advances of machine learning (ML) in computer vision and machine translation, creating learning techniques that are spatially-generalizable and physics-conforming remains an understudied and challenging task in Earth Science (ES). In particular, direct applications of typical ML models often fall short due to two major challenges posed by ES data. First, a fundamental property of spatial data is spatial heterogeneity, which means the functional relationships between target variables (e.g., land cover changes, water temperature and streamflow) and Earth observations tend to be non-stationary over space. The footprints of such heterogeneous data generation functions are often unknown, adding an extra layer of complication. Second, annotated data available in ES applications are often limited or highly localized due to the substantial human labor and material cost for data collection. As a result, pure data-driven attempts – which are often carried out without consideration of underlying physics – are known to be susceptible in learning spurious patterns that overfit limited training data and cannot generalize to large and diverse regions. We aim to explore new model-agnostic learning frameworks to explicitly incorporate spatial heterogeneity awareness and physical knowledge to tackle these challenges in ES. First, to harness spatial heterogeneity, we will explore a statistically-guided framework to automatically capture spatial footprints of data generated by different functions (e.g., target predictions as functions of spectral bands) and transform an user-selected deep network architecture into a heterogeneity-aware version. We will also investigate more effective spatial knowledge sharing models using the spatial-heterogeneity-aware architecture. Second, to further improve the interpretability and generalizability for data-sparse regions, we will explore new physics-guided ML architectures to incorporate domain knowledge, e.g., water temperature dynamics driven by the heat transfer process and other general physical processes embedded in physics-based models. To address the biased parameterizations of physics-based models, we will also investigate new learning strategies to effectively extract general physical relations from multiple physics-based models. Finally, we will explore synergistic integration of the statistically and physically guided frameworks to create a more holistic solution to address various challenging ES application scenarios. Results will be evaluated using important ES tasks, including land cover and land use change (LCLUC) mapping and surface water monitoring, with support from chief scientists of NASA LCLUC and USGS water monitoring programs. To improve the confidence in the success of the new technology, our preliminary exploration has created prototypes of the frameworks to perform statistically-guided spatial transformation for data with spatial heterogeneity, and physics-guided learning for scenarios with limited data. Preliminary case studies using ES data have demonstrated feasibility and the potential of the new frameworks: (1) for land cover mapping, our prototype of spatial transformation improved the F1-score by 10-20% over existing deep learning baselines; and (2) for water temperature and streamflow prediction, the preliminary physics-guided learning model demonstrated improvements over both existing process-based models used by USGS and ML models over large-scale river basins and lakes, and the model has been included in USGS's water prediction workplan. The proposal team includes experts from both computer science and ES. Targeted deliverables of this project include the new technology as well as its open-source implementation, and related ES benchmark datasets used for validation.

Advance Earth system science knowledge through the Identification, develop, and demonstrate innovative information systems technologies

This is early/mid-stage (TRL 5) — the most realistic path in is NASA SBIR/STTR, which funds small businesses and research institutions to develop technology aligned with NASA's needs (equity-free, phased funding). Check whether a current SBIR/STTR solicitation topic overlaps with this project's technology area, or contact the project directly (above) to ask.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.