An Unsupervised Transformer-Based Multivariate Alteration Detection Approach for Change Detection in VHR Remote Sensing Images DOI Creative Commons
Yizhang Lin, Sicong Liu, Yongjie Zheng

et al.

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2024, Volume and Issue: 17, P. 3251 - 3261

Published: Jan. 1, 2024

Multi-temporal change detection (CD) plays a crucial role in the remote sensing application field. In recent years, supervised deep learning methods have shown excellent performance detecting changes very-high-resolution (VHR) images. However, these require large number of labeled samples for training, making process time-consuming and labor-intensive. Unsupervised approaches are more attractive practical applications since they can produce CD map without relying on any ground reference or prior knowledge. this paper, we propose novel unsupervised approach, named Transformer-based Multivariate Alteration Detection (Trans-MAD). It utilizes pre-detection strategy that combines Compressed Change Vector Analysis (C 2 VA) Iteratively Reweighted (IR-MAD) to generate reliable pseudo-training samples. More accurate robust results be achieved by leveraging IR-MAD detect insignificant incorporating attention mechanism model difference similarity between two distant pixels an image. The proposed Trans-MAD approach was validated VHR bi-temporal satellite datasets, obtained experimental demonstrated its superiority comparing with state-of-the-art methods.

Language: Английский

SpectralGPT: Spectral Remote Sensing Foundation Model DOI
Danfeng Hong, Bing Zhang, Xuyang Li

et al.

IEEE Transactions on Pattern Analysis and Machine Intelligence, Journal Year: 2024, Volume and Issue: 46(8), P. 5227 - 5244

Published: April 3, 2024

The foundation model has recently garnered significant attention due to its potential revolutionize the field of visual representation learning in a self-supervised manner. While most models are tailored effectively process RGB images for various tasks, there is noticeable gap research focused on spectral data, which offers valuable information scene understanding, especially remote sensing (RS) applications. To fill this gap, we created first time universal RS model, named SpectralGPT, purpose-built handle using novel 3D generative pretrained transformer (GPT). Compared existing models, SpectralGPT 1) accommodates input with varying sizes, resolutions, series, and regions progressive training fashion, enabling full utilization extensive Big Data; 2) leverages token generation spatial-spectral coupling; 3) captures spectrally sequential patterns via multi-target reconstruction; 4) trains one million images, yielding over 600 parameters. Our evaluation highlights performance improvements signifying substantial advancing Data applications within geoscience across four downstream tasks: single/multi-label classification, semantic segmentation, change detection.

Language: Английский

Citations

254
Decoupled-and-Coupled Networks: Self-Supervised Hyperspectral Image Super-Resolution With Subpixel Fusion DOI
Danfeng Hong, Jing Yao, Chenyu Li

et al.

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2023, Volume and Issue: 61, P. 1 - 12

Published: Jan. 1, 2023

Enormous efforts have been recently made to super-resolve hyperspectral (HS) images with the aid of high spatial resolution multispectral (MS) images. Most prior works usually perform fusion task by means multifarious pixel-level priors. Yet intrinsic effects a large distribution gap between HS-MS data due differences in and spectral are less investigated. The might be caused unknown sensor-specific properties or highly-mixed information within one pixel (due low resolution). To this end, we propose subpixel-level HS super-resolution framework devising novel decoupled-and-coupled network, called DC-Net, progressively fuse from pixel- subpixel-level, image- feature-level. As name suggests, DC-Net first decouples input into common (or cross-sensor) components eliminate before further fusion, then thoroughly blends them model-guided coupled unmixing (CSU) net. More significantly, append self-supervised learning module behind CSU net guaranteeing material consistency enhance detailed appearance restored product. Extensive experimental results show superiority our method both visually quantitatively achieve significant improvement comparison state-of-the-art.

Language: Английский

Citations

115
CasFormer: Cascaded transformers for fusion-aware computational hyperspectral imaging DOI
Chenyu Li, Bing Zhang, Danfeng Hong

et al.

Information Fusion, Journal Year: 2024, Volume and Issue: 108, P. 102408 - 102408

Published: April 6, 2024

Language: Английский

Citations

67
Artificial intelligence for geoscience: Progress, challenges and perspectives DOI Creative Commons
Tianjie Zhao, Sheng Wang,

Chaojun Ouyang

et al.

The Innovation, Journal Year: 2024, Volume and Issue: 5(5), P. 100691 - 100691

Published: Aug. 23, 2024

Public summary•What does AI bring to geoscience? has been accelerating and deepening our understanding of Earth Systems in an unprecedented way, including the atmosphere, lithosphere, hydrosphere, cryosphere, biosphere, anthroposphere interactions between spheres.•What are noteworthy challenges As we embrace huge potential geoscience, several arise reliability interpretability, ethical issues, data security, high demand cost.•What is future The synergy traditional principles modern AI-driven techniques holds immense promise will shape trajectory geoscience upcoming years.AbstractThis paper explores evolution geoscientific inquiry, tracing progression from physics-based models data-driven approaches facilitated by significant advancements artificial intelligence (AI) collection techniques. Traditional models, which grounded physical numerical frameworks, provide robust explanations explicitly reconstructing underlying processes. However, their limitations comprehensively capturing Earth's complexities uncertainties pose optimization real-world applicability. In contrast, contemporary particularly those utilizing machine learning (ML) deep (DL), leverage extensive glean insights without requiring exhaustive theoretical knowledge. ML have shown addressing science-related questions. Nevertheless, such as scarcity, computational demands, privacy concerns, "black-box" nature hinder seamless integration into geoscience. methodologies hybrid presents alternative paradigm. These incorporate domain knowledge guide methodologies, demonstrate enhanced efficiency performance with reduced training requirements. This review provides a comprehensive overview research paradigms, emphasizing untapped opportunities at intersection advanced It examines major showcases advances large-scale discusses prospects that landscape outlines dynamic field ripe possibilities, poised unlock new understandings further advance exploration.Graphical abstract

Language: Английский

Citations

61
Multimodal artificial intelligence foundation models: Unleashing the power of remote sensing big data in earth observation DOI Creative Commons
Danfeng Hong, Chenyu Li, Bing Zhang

et al.

The Innovation Geoscience, Journal Year: 2024, Volume and Issue: 2(1), P. 100055 - 100055

Published: Jan. 1, 2024

Language: Английский

Citations

41
Conventional to Deep Ensemble Methods for Hyperspectral Image Classification: A Comprehensive Survey DOI Creative Commons
Farhan Ullah, Irfan Ullah, Rehan Ullah Khan

et al.

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2024, Volume and Issue: 17, P. 3878 - 3916

Published: Jan. 1, 2024

Hyperspectral image classification has become a hot research topic. HSI been widely used in wide range of real-world application areas due to the in-depth spectral information stored within each pixel. Noticeably, detailed features - i.e., nonlinear correlation between obtained data and correlating object, generate efficient results that are complex for traditional techniques. Deep Learning (DL) recently validated as an influential feature extractor efficiently identifies issues have arisen various computer vision challenges. This motivates using DL Image Classification (HSIC), which shows promising results. survey provides brief description HSIC compares cutting-edge methodologies field. We will first summarize key challenges HSIC, then we discuss superiority DL-ensemble addressing these issues. In this article, divide state-of-the-art with ensemble into features, spatial combined spatial-spectral order comprehensively critically evaluate progress (future directions well) such HSIC. Furthermore, take account involves substantial percentage labeled training images, whereas obtaining number is time cost-consuming. As result, describes some improving performance techniques, can serve future recommendations.

Language: Английский

Citations

40
Coarse to fine-based image–point cloud fusion network for 3D object detection DOI
Meilan Hao, Z.Y. Zhang, P. R. Li

et al.

Information Fusion, Journal Year: 2024, Volume and Issue: 112, P. 102551 - 102551

Published: July 2, 2024

Language: Английский

Citations

30
Vertical accuracy assessment of freely available global DEMs (FABDEM, Copernicus DEM, NASADEM, AW3D30 and SRTM) in flood-prone environments DOI Creative Commons
Michael E. Meadows, Simon Jones, Karin Reinke

et al.

International Journal of Digital Earth, Journal Year: 2024, Volume and Issue: 17(1)

Published: Jan. 25, 2024

Flood models rely on accurate topographic data representing the bare earth ground surface. In many parts of world, only available are free, satellite-derived global Digital Elevation Models (DEMs). However, these have well-known inaccuracies due to limitations sensors used generate them (such as a failure fully penetrate vegetation canopies and buildings). We assess five contemporary, 1 arc-second (≈30 m) DEMs -- FABDEM, Copernicus DEM, NASADEM, AW3D30 SRTM using diverse reference dataset comprised 65 airborne-LiDAR surveys, selected represent biophysical variations in flood-prone areas globally. While vertical accuracy is nuanced, contingent specific metrics character site being assessed, we found that recently-released FABDEM consistently ranked first, improving second-place DEM by reducing large positive errors associated with forests buildings. Our results suggest land cover main factor explaining (especially forests), steep slopes wider error spreads (although resampled from higher-resolution products less sensitive), variable dependency terrain aspect likely function horizontal geolocation problematic for DEM).

Language: Английский

Citations

28
Remote Sensing for Agriculture in the Era of Industry 5.0—A Survey DOI Creative Commons
Nancy Victor, Praveen Kumar Reddy Maddikunta, Delphin Raj Kesari Mary

et al.

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Journal Year: 2024, Volume and Issue: 17, P. 5920 - 5945

Published: Jan. 1, 2024

Agriculture can be regarded as the backbone of human civilization. As technology evolved, synergy between agriculture and remote sensing has brought about a paradigm shift, thereby entirely revolutionizing traditional agricultural practices. Nevertheless, adoption technologies in face various challenges terms limited spatial temporal coverage, high cloud cover, low data quality so on. Industry 5.0 marks new era industrial revolution, where humans machines collaborate closely, leveraging their distinct capabilities, enhancing decision making sustainability resilience. This paper provides comprehensive survey on related aspects dealing with practices (I5.0) era. We also elaborately discuss applications pertaining to I5.0- enabled for agriculture. Finally, we several issues integration I5.0 sensing. offers valuable insights into current state, challenges, potential advancements principles agriculture, thus paving way future research, development, implementation strategies this domain.

Language: Английский

Citations

25
UANet: An Uncertainty-Aware Network for Building Extraction From Remote Sensing Images DOI
Jiepan Li, Wei He, Weinan Cao

et al.

IEEE Transactions on Geoscience and Remote Sensing, Journal Year: 2024, Volume and Issue: 62, P. 1 - 13

Published: Jan. 1, 2024

Building extraction aims to segment building pixels from remote sensing images and plays an essential role in many applications, such as city planning urban dynamic monitoring. Over the past few years, deep learning methods with encoder–decoder architectures have achieved remarkable performance due their powerful feature representation capability. Nevertheless, varying scales styles of buildings, conventional models always suffer uncertain predictions cannot accurately distinguish complete footprints complex distribution ground objects, leading a large degree omission commission. In this paper, we realize importance prediction propose novel straightforward Uncertainty-Aware Network (UANet) alleviate problem. Specifically, first apply general network obtain map relatively high uncertainty. Second, order aggregate useful information highest-level features, design Prior Information Guide Module guide features prior map. Third, based on map, introduce Uncertainty Rank Algorithm measure uncertainty level each pixel belonging foreground background. We further combine algorithm proposed Fusion facilitate level-by-level refinement final refined low To verify our UANet, conduct extensive experiments three public datasets, including WHU dataset, Massachusetts Inria aerial image dataset. Results demonstrate that UANet outperforms other state-of-the-art algorithms by margin. The source code is available at https://github.com/Henryjiepanli/Uncertainty-aware-Network.

Language: Английский

Citations

24