TY - JOUR
T1 - Geoscience-aware deep learning
T2 - A new paradigm for remote sensing
AU - Ge, Yong
AU - Zhang, Xining
AU - Atkinson, Peter M.
AU - Stein, Alfred
AU - Li, Lianfa
N1 - Publisher Copyright:
© 2022 The Authors
PY - 2022/6
Y1 - 2022/6
N2 - Information extraction is a key activity for remote sensing images. A common distinction exists between knowledge-driven and data-driven methods. Knowledge-driven methods have advanced reasoning ability and interpretability, but have difficulty in handling complicated tasks since prior knowledge is usually limited when facing the highly complex spatial patterns and geoscience phenomena found in reality. Data-driven models, especially those emerging in machine learning (ML) and deep learning (DL), have achieved substantial progress in geoscience and remote sensing applications. Although DL models have powerful feature learning and representation capabilities, traditional DL has inherent problems including working as a black box and generally requiring a large number of labeled training data. The focus of this paper is on methods that integrate domain knowledge, such as geoscience knowledge and geoscience features (GK/GFs), into the design of DL models. The paper introduces the new paradigm of geoscience-aware deep learning (GADL), in which GK/GFs and DL models are combined deeply to extract information from remote sensing data. It first provides a comprehensive summary of GK/GFs used in GADL, which forms the basis for subsequent integration of GK/GFs with DL models. This is followed by a taxonomy of approaches for integrating GK/GFs with DL models. Several approaches are detailed using illustrative examples. Challenges and research prospects in GADL are then discussed. Developing more novel and advanced methods in GADL is expected to become the prevailing trend in advancing remotely sensed information extraction in the future.
AB - Information extraction is a key activity for remote sensing images. A common distinction exists between knowledge-driven and data-driven methods. Knowledge-driven methods have advanced reasoning ability and interpretability, but have difficulty in handling complicated tasks since prior knowledge is usually limited when facing the highly complex spatial patterns and geoscience phenomena found in reality. Data-driven models, especially those emerging in machine learning (ML) and deep learning (DL), have achieved substantial progress in geoscience and remote sensing applications. Although DL models have powerful feature learning and representation capabilities, traditional DL has inherent problems including working as a black box and generally requiring a large number of labeled training data. The focus of this paper is on methods that integrate domain knowledge, such as geoscience knowledge and geoscience features (GK/GFs), into the design of DL models. The paper introduces the new paradigm of geoscience-aware deep learning (GADL), in which GK/GFs and DL models are combined deeply to extract information from remote sensing data. It first provides a comprehensive summary of GK/GFs used in GADL, which forms the basis for subsequent integration of GK/GFs with DL models. This is followed by a taxonomy of approaches for integrating GK/GFs with DL models. Several approaches are detailed using illustrative examples. Challenges and research prospects in GADL are then discussed. Developing more novel and advanced methods in GADL is expected to become the prevailing trend in advancing remotely sensed information extraction in the future.
KW - Data-driven
KW - Deep learning
KW - Geoscience features
KW - Geoscience knowledge
KW - Information extraction
KW - Knowledge-driven
KW - ITC-GOLD
U2 - 10.1016/j.srs.2022.100047
DO - 10.1016/j.srs.2022.100047
M3 - Article
AN - SCOPUS:85136137096
SN - 2666-0172
VL - 5
JO - Science of Remote Sensing
JF - Science of Remote Sensing
M1 - 100047
ER -