Scholarlink

<p>Arbitrary-scale super-resolution (SR) for remote sensing is challenging because reconstruction quality depends not only on the upsampling module itself, but also on scale-dependent low-resolution synthesis, heterogeneous multimodal cues, and stability across neighboring target scales. We present ASSR, a multimodal arbitrary-scale SR framework for Sentinel-1/Sentinel-2 reconstruction under a reduced-resolution protocol. ASSR uses scale-conditioned degradation to generate supervision tailored to different target magnifications, dynamic gated fusion to selectively inject Sentinel-1 structural cues into Sentinel-2 features, and an LR-only semantic prior to modulate fusion behavior at the scene level. A compact meta-upsampler allows prediction at arbitrary scales between <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1.5\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1.5</mn> <mo>×</mo> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(6\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>6</mn> <mo>×</mo> </mrow> </math></EquationSource> </InlineEquation>, while a cross-scale consistency objective regularizes neighboring predictions. Quantitative experiments on SEN12MS show consistent improvements in reconstruction fidelity and structural preservation over strong fixed-scale, multimodal, and arbitrary-scale baselines. Zero-shot evaluation on BigEarthNet-MM under the same reduced-resolution setting further suggests that the design transfers reasonably beyond the training benchmark. Additional ablations support the roles of scale-conditioned supervision, gated multimodal fusion, LR semantic modulation, and cross-scale consistency. Overall, ASSR provides a lightweight and effective design for multimodal arbitrary-scale SR in remote sensing, with empirical gains consistently demonstrated under the adopted reduced-resolution protocol.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal arbitrary-scale super-resolution via dynamic gated fusion and low-resolution semantic guidance

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Haibo Jin,
  • Yuxuan Deng

摘要

Arbitrary-scale super-resolution (SR) for remote sensing is challenging because reconstruction quality depends not only on the upsampling module itself, but also on scale-dependent low-resolution synthesis, heterogeneous multimodal cues, and stability across neighboring target scales. We present ASSR, a multimodal arbitrary-scale SR framework for Sentinel-1/Sentinel-2 reconstruction under a reduced-resolution protocol. ASSR uses scale-conditioned degradation to generate supervision tailored to different target magnifications, dynamic gated fusion to selectively inject Sentinel-1 structural cues into Sentinel-2 features, and an LR-only semantic prior to modulate fusion behavior at the scene level. A compact meta-upsampler allows prediction at arbitrary scales between \(1.5\times \) 1.5 × and \(6\times \) 6 × , while a cross-scale consistency objective regularizes neighboring predictions. Quantitative experiments on SEN12MS show consistent improvements in reconstruction fidelity and structural preservation over strong fixed-scale, multimodal, and arbitrary-scale baselines. Zero-shot evaluation on BigEarthNet-MM under the same reduced-resolution setting further suggests that the design transfers reasonably beyond the training benchmark. Additional ablations support the roles of scale-conditioned supervision, gated multimodal fusion, LR semantic modulation, and cross-scale consistency. Overall, ASSR provides a lightweight and effective design for multimodal arbitrary-scale SR in remote sensing, with empirical gains consistently demonstrated under the adopted reduced-resolution protocol.