A survey of machine learning and deep learning in remote sensing of geological environment: Challenges, advances, and opportunities

Due to limited resources and environmental pollution, monitoring the geological environment has become essential for many countries’ sustainable development. As various high-resolution remote-sensing (RS) imaging platforms are continuously available, the remote sensing of the geological environment (GERS) provides a fine-grain, all-weather, and low-cost method for identifying geological elements. Mainstream machine learning (ML) and deep learning (DL) methods can extract high-level high-dimensional semantic information and thus supply an efficient tool for high-precision classification and recognition in many fields. Therefore, the integration of advanced methods and multi-source RS images for GERS interpretation has achieved remarkable breakthroughs during the past decades. However, to the best of our knowledge, a systematic survey of the advances of GERS interpretation regarding ML and DL methods is still lacking. Through the collection of extensive published research in this area, this survey outlines and analyzes the challenges, progress, and promising directions of GERS interpretation. Specifically, the main challenges and difficulties in identifying GERS elements are first summarized in four aspects: sufficient element characteristics and variations, complex context disturbance, RS image quality and types, and other limitations in GERS interpretation. Second, we systematically introduce various RS imaging platforms and advanced ML and DL methods for GERS interpretation. Third, the research status and trends of several GERS applications, including their use for lithology, soil, water, rock glacier, and geological disaster, are ultimately collected and compared. Finally, potential opportunities for future research are discussed. After the systematic and comprehensive review, the conclusive findings suggest that longtime large-scale GERS interpretation and corresponding change pattern analysis will be a significant future direction to meet the needs of environment improvement and sustainable development. To complete the above goals, a fusion of satellite, airplane, environmental monitoring, geological survey, and other types of data will provide enough discriminative information, and expert knowledge, GIS, and high-performance computing techniques will be helpful to improve the efficiency and generalizability of ML and DL methods for processing the multi-platform RS data.