Scholarlink

<p>Remote sensing spatiotemporal fusion is an effective method for generating high spatiotemporal resolution remote sensing images. Current spatiotemporal fusion models use a single pair of coarse-fine resolution images for prediction. They struggle to capture significant changes at the same location over short time intervals. This limitation makes it difficult to achieve high-accuracy image predictions. However, by using two pairs of coarse-fine resolution images for auxiliary prediction, continuous changes in data can be fully utilized, effectively addressing the issue of dramatic image changes. Therefore, this paper proposes the Two-Stream High Temporal Sensitive Convolutional Neural Network (TSCNN). TSCNN introduces the Long-Term Dynamic Capture module (LTDC), which incorporates the temporal variation information extracted from coarse-resolution images into fine-resolution images, enhancing the weight of post-prediction images and preserving the integrity of temporal information, thus reducing the uncertainty of image mutations. To address the issue of inconsistent multiscale features in remote sensing images and improve image accuracy, we design an encoder-decoder structure. The encoder uses a multiscale extraction module to extract features from remote sensing images, expanding the range of contextual information and enhancing the utilization of features at different scales. In the decoder, the Global-Local Attentional Feature Fusion module (GLAFF) is introduced to capture global and local semantic features, which are then fused and reconstructed. This approach improves the completeness of the acquired information. Based on the quantitative results of three datasets, the proposed method improves the quantitative indices of spectral angle mapper (SAM), relative dimensionless global error (ERGAS), correlation coefficient (CC), and structural similarity (SSIM) by up to 0.3360, 0.1111, 0.0308, and 0.0308 at most compared with the best comparison method. TSCNN improves the accuracy of prediction. It has a wide application prospect.</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

A spatiotemporal fusion method based on two-stream high temporal sensitive convolutional neural network

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Yujia Li,
  • Dajiang Lei,
  • Qianwei Zhu,
  • Junmin Wang,
  • Liping Zhang,
  • Weisheng Li

摘要

Remote sensing spatiotemporal fusion is an effective method for generating high spatiotemporal resolution remote sensing images. Current spatiotemporal fusion models use a single pair of coarse-fine resolution images for prediction. They struggle to capture significant changes at the same location over short time intervals. This limitation makes it difficult to achieve high-accuracy image predictions. However, by using two pairs of coarse-fine resolution images for auxiliary prediction, continuous changes in data can be fully utilized, effectively addressing the issue of dramatic image changes. Therefore, this paper proposes the Two-Stream High Temporal Sensitive Convolutional Neural Network (TSCNN). TSCNN introduces the Long-Term Dynamic Capture module (LTDC), which incorporates the temporal variation information extracted from coarse-resolution images into fine-resolution images, enhancing the weight of post-prediction images and preserving the integrity of temporal information, thus reducing the uncertainty of image mutations. To address the issue of inconsistent multiscale features in remote sensing images and improve image accuracy, we design an encoder-decoder structure. The encoder uses a multiscale extraction module to extract features from remote sensing images, expanding the range of contextual information and enhancing the utilization of features at different scales. In the decoder, the Global-Local Attentional Feature Fusion module (GLAFF) is introduced to capture global and local semantic features, which are then fused and reconstructed. This approach improves the completeness of the acquired information. Based on the quantitative results of three datasets, the proposed method improves the quantitative indices of spectral angle mapper (SAM), relative dimensionless global error (ERGAS), correlation coefficient (CC), and structural similarity (SSIM) by up to 0.3360, 0.1111, 0.0308, and 0.0308 at most compared with the best comparison method. TSCNN improves the accuracy of prediction. It has a wide application prospect.