Title: Spatially-explicit and knowledge-guided AI for Earth Science
Presenting Author: Yiqun Xie
Organization: University of Maryland
Co-Author(s): Xiaowei Jia, Sergii Skakun

Abstract:
The generalizability of machine learning algorithms over space and time has been a major challenge in Earth science applications, due to the variability over geographic regions, and limited ground truth observations both in terms of quantity and spatial coverage. We aim to tackle these challenges with: (1) New spatially-explicit learning frameworks, which can automatically recognize regional variability, and create local models to avoid confusion and enhance prediction quality. The framework is model-agnostic and can be used with both deep learning and traditional learning methods such as random forest. Our results for crop mapping in CONUS showed the effectiveness of the framework in capturing spatial variability and significantly improving accuracy over many areas. (2) Knowledge-guided learning frameworks, where prior knowledge about physics or spatio-temporal patterns can be incorporated into deep learning models to provide valuable extra supervision and reduce the difficulty of learning. The knowledge can also be leveraged to create self-supervised learning strategies to minimize the need of manual annotations for training. Our evaluations showed promising results in hydrological monitoring and satellite imagery cloud masking. Applicability-wise, the spatially-explicit learning framework is suitable for broad applications over large areas that utilize Earth observation datasets for classification, prediction, etc. The knowledge-guided learning methods are more suitable for domains (e.g., hydrology, energy, carbon) where theory-based models have been developed, or problems where knowledge-based rules can be derived.