F-Segfomer: A Feature-Selection Approach for Land Resource Management on Unseen Domains DOI Open Access
Mạnh Hùng Nguyễn, Chi Cuong Vu

Sustainability, Journal Year: 2025, Volume and Issue: 17(6), P. 2640 - 2640

Published: March 17, 2025

Satellite imagery segmentation is essential for effective land resource management. However, diverse geographical landscapes may limit accuracy in practical applications. To address these challenges, we propose the F-Segformer network, which incorporates a Variational Information Bottleneck (VIB) module to enhance feature selection within SegFormer architecture. The VIB serves as selector, providing improved regularization, while well adapted unseen domains. Combining methods, our robustly enhanced performance new regions that do not appear training process. Additionally, employ Online Hard Example Mining (OHEM) prioritize challenging samples during training, setting helps with accelerating model convergence even co-trained loss. Experimental results on LoveDA dataset show method can achieve comparable result well-known domain-adaptation methods without using data from target domain. In scenario when trained domain and tested an domain, shows significant improvement. Last but least, OHME converge three times faster than OHME.

Language: Английский

U-Net-Based Medical Image Segmentation DOI Creative Commons

Xiaoxia Yin,

Le Sun,

Yuhan Fu

et al.

Journal of Healthcare Engineering, Journal Year: 2022, Volume and Issue: 2022, P. 1 - 16

Published: April 15, 2022

Deep learning has been extensively applied to segmentation in medical imaging. U-Net proposed 2015 shows the advantages of accurate small targets and its scalable network architecture. With increasing requirements for performance imaging recent years, cited academically more than 2500 times. Many scholars have constantly developing This paper summarizes image technologies based on structure variants concerning their structure, innovation, efficiency, etc.; reviews categorizes related methodology; introduces loss functions, evaluation parameters, modules commonly imaging, which will provide a good reference future research.

Language: Английский

Citations

238
Improving agricultural field parcel delineation with a dual branch spatiotemporal fusion network by integrating multimodal satellite data DOI

Zhiwen Cai,

Qiong Hu, Xinyu Zhang

et al.

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 205, P. 34 - 49

Published: Oct. 5, 2023

Language: Английский

Citations

33
FCD-AttResU-Net: An improved forest change detection in Sentinel-2 satellite images using attention residual U-Net DOI Creative Commons
Kassim Kalinaki, Owais Ahmed Malik,

Daphne Teck Ching Lai

et al.

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2023, Volume and Issue: 122, P. 103453 - 103453

Published: Aug. 1, 2023

Forest Change Detection (FCD) is a critical component of natural resource monitoring and conservation strategies, enabling informed decision-making. Various methods utilizing the power artificial intelligence (AI) have been developed for detecting categorizing changes in forest cover using remote sensing (RS) data. One prominent AI-powered approach U-Net, deep learning (DL) architecture famous its segmentation proficiency. However, standard U-Net fails to effectively capture intricate spatial dependencies long-range contextual information present imagery. To address this research gap, we introduce an attention-residual-based novel DL model which leverages Sentinel-2 satellite images map alterations vegetation tropical region. Our enhances by seamlessly integrating strengths harnessing attention mechanisms strategically amplify crucial features, leveraging cutting-edge residual connections facilitate smooth flow gradient propagation. These meticulous design choices enabled precise feature extraction, resulting improved computational performance proposed method compared Standard Deeplabv3+, Deep Res-U-Net, Attention U-Net. The classification results demonstrate enhanced efficiency our model, achieving Mean Intersection over Union (MIoU) 0.9330 on test dataset. This surpasses (0.9146), (0.9029), Deeplabv3+ (0.9247), Res-U-Net (0.9282). comparative analysis ground truth reproductions unveiled superior detection capabilities accurately identifying non-forest polygons, surpassing both augmented with mechanism, along other state-of-the-art techniques, thereby highlighting efficacy. model's broad applicability can support managers ecologists rapidly evaluating long-term ramifications infrastructure initiatives, such as roads, forests, including those Brunei.

Language: Английский

Citations

24
TransU-Net++: Rethinking attention gated TransU-Net for deforestation mapping DOI Creative Commons
Ali Jamali, Swalpa Kumar Roy, Jonathan Li

et al.

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2023, Volume and Issue: 120, P. 103332 - 103332

Published: May 11, 2023

Deforestation has become a major cause of climate change, and as result, both characterizing the drivers estimating segmentation maps deforestation have piqued interest researchers. In computer vision domain, Vision Transformers (ViTs) shown their superiority compared to extensively utilized convolutional neural networks (CNNs) over last couple years. Although, ViTs several challenges, specifically in remote sensing image processing, including significant complexity that increases computation costs need for much higher reference data than CNNs. As such, this paper, we introduce an attention gates aided TransU-Net, called TransU-Net++ semantic with application mapping two South American forest biomes, i.e., Atlantic Forest Amazon Rainforest. The heterogeneous kernel convolution (HetConv), U-Net, gates, are all proposed advantage. significantly increased performance TransU-Net's dataset by about 4%, 6%, 16%, respectively, terms overall accuracy, F1-score, recall, respectively.Moreover, results show developed TrasnU-Net++ model (0.921) achieves highest Area under ROC Curve value 3-band other models, ICNet (0.667), ENet (0.69), SegNet (0.788), U-Net (0.871), Attention U-Net-2 (0.886), R2U-Net (0.888), TransU-Net (0.889), Swin (0.893), ResU-Net (0.896), U-Net+++ (0.9), (0.908), respectively. code will be made publicly available at https://github.com/aj1365/TransUNetplus2.

Language: Английский

Citations

23
Harnessing Time-Series Satellite Data and Deep Learning to Monitor Historical Patterns of Deforestation in Eastern Himalayan Foothills of India DOI
Jintu Moni Bhuyan, Subrata Nandy, Hitendra Padalia

et al.

Journal of the Indian Society of Remote Sensing, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 19, 2025

Language: Английский

Citations

1
Improved U-Net Remote Sensing Classification Algorithm Based on Multi-Feature Fusion Perception DOI Creative Commons
Chuan Yan, Xiangsuo Fan,

Jinlong Fan

et al.

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(5), P. 1118 - 1118

Published: Feb. 24, 2022

The selection and representation of remote sensing image classification features play crucial roles in accuracy. To effectively improve the accuracy features, an improved U-Net network framework based on multi-feature fusion perception is proposed this paper. This adds channel attention module (CAM-UNet) to original cascades shallow with deep semantic replaces layer a support vector machine, finally uses majority voting game theory algorithm fuse multifeature results obtain final results. study used forest distribution Xingbin District, Laibin City, Guangxi Zhuang Autonomous Region as research object, which Landsat 8 multispectral images, and, by combining spectral spatial advanced overcame influence reduction resolution that occurs deepening experimental showed can Before improvement, overall segmentation forestland increased from 90.50% 92.82% 95.66% 97.16%, respectively. cover obtained paper be input data for regional ecological models, conducive development accurate real-time vegetation growth change models.

Language: Английский

Citations

36
FCD-R2U-net: Forest change detection in bi-temporal satellite images using the recurrent residual-based U-net DOI
Ehsan Khankeshizadeh, Ali Mohammadzadeh, Armin Moghimi

et al.

Earth Science Informatics, Journal Year: 2022, Volume and Issue: 15(4), P. 2335 - 2347

Published: Nov. 2, 2022

Language: Английский

Citations

33
High-Resolution Semantic Segmentation of Woodland Fires Using Residual Attention UNet and Time Series of Sentinel-2 DOI Creative Commons
Zeinab Shirvani, Omid Abdi,

Rosa C. Goodman

et al.

Remote Sensing, Journal Year: 2023, Volume and Issue: 15(5), P. 1342 - 1342

Published: Feb. 28, 2023

Southern Africa experiences a great number of wildfires, but the dependence on low-resolution products to detect and quantify fires means both that there is time lag many small fire events are never identified. This particularly relevant in miombo woodlands, where frequent predominantly small. We developed cutting-edge deep-learning-based approach uses freely available Sentinel-2 data for near-real-time, high-resolution detection Mozambique. The importance main bands their derivatives was evaluated using TreeNet, top five variables were selected create three training datasets. designed UNet architecture, including contraction expansion paths bridge between them with several layers functions. then added attention gate units (AUNet) residual blocks (RAUNet) architecture. trained models efficiency all high (intersection over union (IoU) > 0.85) increased more variables. first an RAUNet architecture has been used events, it performed better than AUNet models—especially detecting fires. model had IoU = 0.9238 overall accuracy 0.985. suggest others test large datasets from different regions other satellites so may be applied broadly improve wildfires.

Language: Английский

Citations

22
Deep-Learning for Change Detection Using Multi-Modal Fusion of Remote Sensing Images: A Review DOI Creative Commons

Souad Saidi,

Soufiane Idbraim,

Younes Karmoude

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(20), P. 3852 - 3852

Published: Oct. 17, 2024

Remote sensing images provide a valuable way to observe the Earth’s surface and identify objects from satellite or airborne perspective. Researchers can gain more comprehensive understanding of by using variety heterogeneous data sources, including multispectral, hyperspectral, radar, multitemporal imagery. This abundance different information over specified area offers an opportunity significantly improve change detection tasks merging fusing these sources. review explores application deep learning for in remote imagery, encompassing both homogeneous scenes. It delves into publicly available datasets specifically designed this task, analyzes selected models employed detection, current challenges trends field, concluding with look towards potential future developments.

Language: Английский

Citations

7
Improved U-Net Remote Sensing Classification Algorithm Fusing Attention and Multiscale Features DOI Creative Commons
Xiangsuo Fan, Chuan Yan,

Jinlong Fan

et al.

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(15), P. 3591 - 3591

Published: July 27, 2022

The selection and representation of classification features in remote sensing image play crucial roles accuracy. To effectively improve the accuracy, an improved U-Net algorithm fusing attention multiscale is proposed this paper, called spatial attention-atrous pyramid pooling (SA-UNet). This framework connects atrous (ASPP) with convolutional units encoder original form residuals. ASPP module expands receptive field, integrates network, enhances ability to express shallow features. Through fusion residual module, deep are deeply fused, characteristics further used. mechanism used combine semantic information so that decoder can recover more information. In study, crop distribution central Guangxi province was analyzed, experiments were conducted based on Landsat 8 multispectral images. experimental results showed increases accuracy increasing from 93.33% 96.25%, segmentation sugarcane, rice, other land increased 96.42%, 63.37%, 88.43% 98.01%, 83.21%, 95.71%, respectively. agricultural planting area obtained by be as input data for regional ecological models, which conducive development accurate real-time growth change models.

Language: Английский

Citations

27