Adapting Semi-Supervised Segmentation methods to Multimodal Remote Sensing Data

计算机科学 分割 人工智能 计算机视觉 模式识别(心理学) 遥感 地理
作者
Itza Hernandez-Sequeira,Damian Ibañez,Rubén Fernández-Beltrán,Filiberto Pla
出处
期刊:The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences [Copernicus Publications]
卷期号:XLVIII-M-7-2025: 21-28
标识
DOI:10.5194/isprs-archives-xlviii-m-7-2025-21-2025
摘要

Abstract. Remote sensing (RS) imagery is important for applications ranging from land cover and land use (LCLU) mapping to agriculture and forest monitoring. However, there is a limited availability of high-quality labeled data to use as a reference to train supervised learning (SL) models. Semi-supervised learning (SSL) frameworks, such as UniMatch (Yang et al., 2023), use pseudo-labeling and consistency regularization methods to address this limitation. Similar works have been adapted to RS: LSST (Lu et al., 2022) refines pseudo-labels with adaptive class-specific thresholds, while RS-DWL (Huang et al., 2024) mitigates noise and class imbalance through decoupled learning and confidence-based weighting. Despite these advances, SSL applications to multimodal RS imagery remain underexplored. We address this gap by adapting the SSL framework UniMatch to incorporate diverse encoders and multimodal remote sensing data for LCLU segmentation. We experimented on FLAIR-2 (Garioud et al., 2023), a dataset that combines very high-resolution aerial imagery (RGB) with near-infrared (NIR) data and elevation measurements (above-ground height). Key findings reveal that we achieved the best segmentation results using a transformer encoder for SL and SSL scenarios. When comparing RGB-only data and multimodal data, we observed that some classes, like “buildings”, “water”, and “coniferous”, benefited from the inclusion of NIR and elevation information. In the semi-supervised experiments, where only half of the data was labeled, and the remaining half was used as unlabeled (simulating a real-world scenario), the multimodal SSL approach outperformed the fully supervised learning (FSL) approach using only the labeled subset (1/2). These results highlight the strong potential of data fusion in RS applications with limited labeled data.

科研通智能强力驱动
Strongly Powered by AbleSci AI
科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
刚刚
刚刚
Owen的应助被pepsi191采纳,获得10
2秒前
3秒前
3秒前
4秒前
可爱的函函的应助被FiiisH采纳,获得10
4秒前
111完成签到,获得积分10
5秒前
李健的小迷弟的应助被小涵采纳,获得10
5秒前
6秒前
Akim的应助被dbwdqqr采纳,获得10
6秒前
tlc_191026发布了新的文献求助10
6秒前
李爱国的应助被renn采纳,获得20
7秒前
7秒前
7秒前
9秒前
xszlll发布了新的文献求助10
10秒前
11秒前
11秒前
chengyu完成签到,获得积分10
14秒前
14秒前
pepsi191发布了新的文献求助10
14秒前
15秒前
btbu2015发布了新的文献求助10
15秒前
郑睿涛发布了新的文献求助10
15秒前
江以轩完成签到 ,获得积分10
16秒前
17秒前
王昕钥完成签到,获得积分10
17秒前
OLIVIA完成签到,获得积分10
17秒前
烟花的应助被salura采纳,获得10
17秒前
科研通AI6.3的应助被简单生活采纳,获得10
18秒前
777发布了新的文献求助10
18秒前
tlc_191026发布了新的文献求助10
18秒前
19秒前
molihuakai的应助被激动的绿竹采纳,获得10
20秒前
dbwdqqr发布了新的文献求助10
20秒前
MiyaGuo发布了新的文献求助10
20秒前
lfp发布了新的文献求助10
24秒前
小蘑菇的应助被丹皮小姐采纳,获得10
25秒前
CipherSage的应助被renn采纳,获得20
28秒前
高分求助中
(应助此贴封号)【重要!!请各用户(尤其是新用户)详细阅读】【科研通的精品贴汇总】 10000
拉伸高分子熔体中缠结行为的模拟和小角中子散射研究 1700
The World Atlas of Wine 9th edition 1000
奈特基础组织学:配套组织病理学(第4版)(Netter's Essential Histology With Correlated Histopathology) 800
八旗制度の研究 600
Improving the wear resistance of vehicle gear oil by oil soluble LaF3 nanoparticles and typical additives of gear oil 500
Boundaries and Bodhisattvas 500
热门求助领域 (近24小时)
化学 材料科学 医学 生物 纳米技术 计算机科学 工程类 有机化学 化学工程 生物化学 物理 复合材料 内科学 催化作用 光电子学 人工智能 细胞生物学 无机化学 物理化学 数学
热门帖子
关注 科研通微信公众号,转发送积分 8183449
求助须知:如何正确求助?哪些是违规求助? 9572834
关于积分的说明 21827509
捐赠科研通 7780646
什么是DOI,文献DOI怎么找? 3434016
关于科研通互助平台的介绍 2558189
邀请新用户注册赠送积分活动 2480371