Overleaf Project Link

© Sakura, 2024. All rights reserved.

This document is protected by copyright law. No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author, except in the case of brief quotations embodied in critical reviews and certain other noncommercial uses permitted by copyright law. For permission requests, please contact the author at [email protected].

Abstract

The advanced techniques of artificial intelligence (AI) in geoscience have revolutionized scientific researches. This survey provides a comprehensive overview of AI's transformative impact on geoscience, including remote sensing, climatology, oceanology, geology and semismology, especially for advanced deep learning methods.

Keywords: Artificial Intelligence, Geoscience, Remote Sensing, Climatology, Oceanology, Geology, Seismology, Physics-Informed Neural Networks, Machine Learning.

• Introduction
• Remote Sensing
• Climatology
• Oceanology
• Geology & Geophysics
• Physics-Informed Neural Networks for GeoScience
• Supplyment
• References

The integration of artificial intelligence (AI) into geoscience marks a paradigm shift in how we analyze, interpret, and predict complex environmental phenomena. Geoscience, a field encompassing disciplines such as remote sensing, climatology, oceanology, and geology, has traditionally relied on extensive data collection and physical models to understand the Earth's systems. However, the sheer volume and complexity of geospatial data present significant challenges that can be effectively addressed through AI's advanced computational capabilities.

Remote sensing, which involves the acquisition of information about the Earth's surface without direct contact, has significantly benefited from AI techniques. Traditional methods of processing remote sensing data, such as manual interpretation and classical statistical approaches, are often time-consuming and prone to errors. AI models, particularly deep learning frameworks, have revolutionized tasks such as image classification, change detection, and semantic segmentation. Techniques like Hierarchical Attention Transformers (HAT), Generative Adversarial Networks (GAN), and UNet-based architectures have set new benchmarks in performance, providing more accurate and reliable data interpretations.

Climatology, the study of climate and its variations, has also seen substantial advancements through AI. Accurate weather forecasting, critical for disaster preparedness and resource management, has been significantly improved by AI-based models. Systems like Pangu-Weather and Aurora leverage deep learning and extensive datasets to predict weather patterns with unprecedented accuracy. These models process vast amounts of historical climate data to generate predictions that are crucial for both short-term weather forecasts and long-term climate change projections.

In oceanology, AI applications have enhanced our understanding and prediction of oceanic phenomena, such as sea ice concentration. Predictive models like SICNet90 utilize deep learning to analyze reanalysis data, offering more precise forecasts of sea ice coverage. The integration of high-quality reanalysis data is vital for improving the accuracy of these predictions, which are essential for navigation, climate studies, and ecological monitoring.

Geology and geophysics, particularly the study of seismic activities, have also been transformed by AI. Seismology foundation models like SeisCLIP and the Seismic Foundation Model (SFM) are pre-trained on multimodal data, enhancing their ability to extract seismic features and perform inversion tasks. These models surpass traditional methods in accuracy and efficiency, offering new insights into earthquake prediction and seismic hazard assessment.

Furthermore, the development of physics-informed neural networks (PINNs) represents a significant innovation in geoscience. PINNs integrate machine learning with physical models, ensuring that predictions adhere to known physical laws. This approach enhances the interpretability and robustness of AI models, making them more reliable for practical applications in geoscience.

This survey aims to provide a comprehensive overview of the current state of AI applications in geoscience. By examining the advancements in model architecture, data integration, and training methodologies, we highlight the transformative impact of AI on this field. The continuous evolution of AI technologies promises to further enhance the accuracy, efficiency, and scope of geoscientific research, offering innovative solutions to complex environmental challenges. As we delve into the various subfields of geoscience, this survey will underscore the critical role of AI in advancing our understanding of Earth's systems and addressing pressing environmental issues.

Classification

Change Detection

Semantic Segmentation

Visual Question Answering

In 2020, Kaplan et al. (the OpenAI team) firstly proposed to model the power-law relationship of model performance with respective to three major factors, namely model size (N), dataset size (D), and the amount of training compute (C), for neural language models, and L(·) denotes the cross entropy loss in nats.¹

1. The statement and figure are adapted from Zhao et al. (2023) ’s work entitled A Survey of Large Language Models (arXiv:2303.18223 [cs]).

Note: The following statements are purely written by author (sakura), which can be biased even misleading to some extent.

​ In the field of remote sensing, a variety of data-driven based models have achieved state-of-the-art performance in different downstream tasks. Hierarchical Attention Transformer (HAT), Generative Adversarial Network (GAN) and Diffusion Probabilistic Model are commonly used in super-resolution. UNet-based models achieve superior performance in object detection and semantic segmentation. Moreover, large visual-language pretraining (VLP) models that incorporate with contextual cues enhances the model capabilities compared to vision-only ones. Nevertheless, these SOTA models, designed for specialized downstream tasks, still serve seperately as a so-called task-agnostic feature learner, which lacks general capacities.

​ Recently, the prosperity of ChatGPT has been witnessed by the AI community. A large number of follow-up work is inspired by this “pre-training and fine-tuning” learning paradigm. Pre-trained on an unprecedentedly large dataset and consumed enourmous computer budget, large language models (LLMs) encompass astonishing learning and understanding capability, which shed a light to achieve artificial general intelligence (AGI).

​ As LLMs seems to be considered as general-purpose task solvers in the field of linguistics, researches are inspired to explore a similar model which is also capable to solve numerous remote sensing tasks in a singular model, in other words, a remote sensing foundation model (RSFM). Encouraging work such as SkySense (Guo et al., 2024) and SpectralGPT (Hong et al., 2024) show that scaling model (e.g., scaling model size or data size) is integral to achieve superior performance, as well as to enable models to handle diverse tasks, learn comprehensive features, and generalize well across various datasets and tasks.

Wuhan University

Overview of SkySense Model Architecture (Guo et al., 2024)

Superior Performance on 16 Datasets over 7 Tasks (Guo et al., 2024)

Chinese Academy of Sciences

Huawei Research

Methodology:

• Objective: Develop a weather forecast system using deep learning.

• Input and Output: The system uses reanalysis weather data from a single point in time as input to predict reanalysis weather data at a future point.

• Data Source: ERA5 data with a time resolution of 1 hour.

• Training Data: The training subset spans from 1979 to 2017, providing 341,880 time points per epoch.

• Overfitting Mitigation: Random permutation of sample order at the start of each epoch.

Model Training:

• Networks: Four deep networks trained for lead times of 1 hour, 3 hours, 6 hours, and 24 hours.

• Duration: Each network trained for 100 epochs, approximately 16 days per network on a cluster of 192 NVIDIA Tesla-V100 GPUs.

Optimization Details:

• Optimizer: Adam optimizer.

• Loss Function: Mean Absolute Error (MAE) loss.

• Normalization: Each two-dimensional input field was normalized by subtracting the mean and dividing by the standard deviation, calculated from 1979 to 2017 data.

• Variable Weights: Weights were inversely proportional to early run loss values, aiming for balanced contributions:

• Upper-Air Variables: Z (3.00), Q (0.60), T (1.50), U (0.77), V (0.54).

• Surface Variables: MSLP (1.50), U10 (0.77), V10 (0.66), T2M (3.00).

• Additional weights added to the MAE loss: 1.0 for upper-air, 0.25 for surface variables.

• Batch Size: 192 (one sample per GPU).

• Learning Rate: Started at 0.0005, annealed to 0 using a cosine schedule.

• Overfitting Mitigation: Weight decay of 3 × 10 -6 and ScheduledDropPath with a drop ratio of 0.2.

• Convergence: Models did not fully converge within 100 epochs, suggesting extended training could improve accuracy.

Results:

Accuracy of tested variables at different lead times (1 hour, 3 hours, 6 hours, and 24 hours) was plotted, indicating the performance of the models.

Microsoft Research

1. Climate Reanalyzer. (n.d.). Retrieved 29 May 2024, from https://climatereanalyzer.org/research_tools/monthly_maps/
2. Eyes on the Earth—NASA/JPL. (n.d.). Eyes on the Earth - NASA/JPL. Retrieved 29 May 2024, from https://eyes.nasa.gov/apps/earth

Chinese Academy of Sciences

University of Science and Technology of China

University of Science and Technology of China

Commonly Used Factors:

• Surface Air Temperature (SAT)

• Sea Surface Temperature (SST)

• Geopotential Height at 500 hPa (500 GH)

National Snow and Ice Data Center (NSIDC)

Data Description

• Purpose: Monitoring sea ice extent, trends, model validation.

• Parameters: Sea ice concentration, fractional coverage.

• Sources: Nimbus-7 SMMR, DMSP-F8, -F11, -F13 SSM/I, DMSP-F17 SSMIS.

• Format: NetCDF with browse images and metadata.

File Information

• Naming Convention: NSIDC0051_SEAICE_PS_HXXkm_YYYYMMDD_v2.0_SSS.ext.

• Components: Grid cell size, date, sensor type.

Spatial and Temporal Information

• Coverage: Polar regions, 25 x 25 km grid.

• Resolution: Varies by latitude.

• Temporal Coverage: Daily and monthly, from October 26, 1978, with gaps.

Data Acquisition and Processing

• Background: Brightness temperatures from SMMR, SSM/I, SSMIS.

• Processing: Inter-sensor corrections, land-to-ocean spillover adjustment, residual weather effects correction, manual quality control, data gap filling.

Quality, Errors, and Limitations

• Accuracy: Better in winter, worse in summer/melt regions.

• Limitations: Sensor differences, residual weather effects, land contamination.

Instrumentation

• Sensors: SMMR, SSM/I, SSMIS - descriptions and differences.

Software and Tools

• Formats: NetCDF.

• Tools: Panoply, NCO, scripts for data handling.

Ahmed, N., Rahman, R. M., Adnan, M. S. G., & Ahmed, B. (2021). Dense prediction of label noise for learning building extraction from aerial drone imagery. International Journal of Remote Sensing, 42(23), 8906–8929. https://doi.org/10.1080/01431161.2021.1973685

Asan, B., Akgül, A., Unal, A., Kandemir, M., & Unal, G. (2024, April 4). Calibrating Bayesian UNet++ for Sub-Seasonal Forecasting. https://doi.org/10.48550/arXiv.2403.16612

Bi, K., Xie, L., Zhang, H., Chen, X., Gu, X., & Tian, Q. (2023). Accurate medium-range global weather forecasting with 3D neural networks. Nature, 619(7970), 533–538. https://doi.org/10.1038/s41586-023-06185-3

Bodnar, C., Bruinsma, W. P., Lucic, A., Stanley, M., Brandstetter, J., Garvan, P., Riechert, M., Weyn, J., Dong, H., Vaughan, A., Gupta, J. K., Tambiratnam, K., Archibald, A., Heider, E., Welling, M., Turner, R. E., & Perdikaris, P. (2024). Aurora: A Foundation Model of the Atmosphere (arXiv:2405.13063). http://arxiv.org/abs/2405.13063

DiGirolamo, N., C. L. Parkinson, D. J. Cavalieri, P. Gloersen, and H. J. Zwally. (2022). Sea Ice Concentrations from Nimbus-7 SMMR and DMSP SSM/I-SSMIS Passive Microwave Data, Version 2 [Data Set]. Boulder, Colorado USA. NASA National Snow and Ice Data Center Distributed Active Archive Center. https://doi.org/10.5067/MPYG15WAA4WX. Date Accessed 05-29-2024.

Guo, X., Lao, J., Dang, B., Zhang, Y., Yu, L., Ru, L., Zhong, L., Huang, Z., Wu, K., Hu, D., He, H., Wang, J., Chen, J., Yang, M., Zhang, Y., & Li, Y. (2024). SkySense: A Multi-Modal Remote Sensing Foundation Model Towards Universal Interpretation for Earth Observation Imagery. Computer Vision and Pattern Recognition (CVPR). https://doi.org/10.48550/arXiv.2312.10115

Hong, D., Zhang, B., Li, X., Li, Y., Li, C., Yao, J., Yokoya, N., Li, H., Ghamisi, P., Jia, X., Plaza, A., Gamba, P., Benediktsson, J. A., & Chanussot, J. (2024). SpectralGPT: Spectral Remote Sensing Foundation Model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1–18. https://doi.org/10.1109/TPAMI.2024.3362475

Huang, X., Ren, L., Liu, C., Wang, Y., Yu, H., Schmitt, M., Hansch, R., Sun, X., Huang, H., & Mayer, H. (2022). Urban Building Classification (UBC) – A Dataset for Individual Building Detection and Classification from Satellite Imagery. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 1412–1420. https://doi.org/10.1109/CVPRW56347.2022.00147

Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020). Scaling Laws for Neural Language Models. ArXiv. https://www.semanticscholar.org/paper/Scaling-Laws-for-Neural-Language-Models-Kaplan-McCandlish/e6c561d02500b2596a230b341a8eb8b921ca5bf2

Li, Q., Mou, L., Sun, Y., Hua, Y., Shi, Y., & Zhu, X. X. (2024). A Review of Building Extraction From Remote Sensing Imagery: Geometrical Structures and Semantic Attributes. IEEE Transactions on Geoscience and Remote Sensing, 62, 1–15. https://doi.org/10.1109/TGRS.2024.3369723

Li, L., Carver, R., Lopez-Gomez, I., Sha, F., & Anderson, J. (2024). Generative emulation of weather forecast ensembles with diffusion models. Science Advances, 10(13), eadk4489. https://doi.org/10.1126/sciadv.adk4489

Li, L., Wang, C., Zhang, H., Zhang, B., & Wu, F. (2019). Urban Building Change Detection in SAR Images Using Combined Differential Image and Residual U-Net Network. Remote Sensing, 11(9), 1091. https://doi.org/10.3390/rs11091091

Ren, Y., & Li, X. (2023). Predicting the Daily Sea Ice Concentration on a Subseasonal Scale of the Pan-Arctic During the Melting Season by a Deep Learning Model. IEEE Transactions on Geoscience and Remote Sensing, 61, 1–15. https://doi.org/10.1109/TGRS.2023.3279089

Shen, C., Appling, A. P., Gentine, P., Bandai, T., Gupta, H., Tartakovsky, A., Baity-Jesi, M., Fenicia, F., Kifer, D., Li, L., Liu, X., Ren, W., Zheng, Y., Harman, C. J., Clark, M., Farthing, M., Feng, D., Kumar, P., Aboelyazeed, D., … Lawson, K. (2023). Differentiable modelling to unify machine learning and physical models for geosciences. Nature Reviews Earth & Environment, 4(8), 552–567. https://doi.org/10.1038/s43017-023-00450-9

Sheng, H., Wu, X., Si, X., Li, J., Zhang, S., & Duan, X. (2023). Seismic Foundation Model (SFM): A new generation deep learning model in geophysics (arXiv:2309.02791). arXiv. https://doi.org/10.48550/arXiv.2309.02791

Si, X., Wu, X., Sheng, H., Zhu, J., & Li, Z. (2024). SeisCLIP: A Seismology Foundation Model Pre-Trained by Multimodal Data for Multipurpose Seismic Feature Extraction. IEEE Transactions on Geoscience and Remote Sensing, 62, 1–13. https://doi.org/10.1109/TGRS.2024.3354456

Zhao, D., Lu, J., & Yuan, B. (2024). See, Perceive, and Answer: A Unified Benchmark for High-Resolution Postdisaster Evaluation in Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 62, 1–14. https://doi.org/10.1109/TGRS.2024.3386934

Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Chen, Z., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., … Wen, J.-R. (2023). A Survey of Large Language Models (arXiv:2303.18223). arXiv. https://doi.org/10.48550/arXiv.2303.18223