Publications

Chang, J; Du, W; Zhang, B; Guo, S; Yin, Y; Wang, Z; Xu, TY; Feng, ZY (2025). Based on the Improved EDCSTFN Model, Modis, Landsat 8, and Sentinel-2 Data Were Fused to Obtain 10 m Dense Time Series Images. IEEE ACCESS, 13, 79189-79202.

Abstract
High temporal and spatial resolution Earth observation data are crucial in remote sensing, but it is difficult to acquire images that guarantee high temporal and spatial resolution simultaneously due to satellite, technology and budget constraints. In this paper, time series image data with 10 m resolution are generated by spatio-temporal fusion of Modis, Landsat and Sentinel data, which reduces the temporal resolution to 1-2 days, while the existing EDCSTFN model is improved in order to overcome the problem of difficulty in global information extraction due to convolution limitation. The encoder and residual encoder use multi-scale convolution to capture more information from raw Landsat data and enhance feature extraction. In addition, a channel attention module (SE) is introduced to model the nonlinear relationship across channels, which improves the nonlinear capability of the model and reduces the sensitivity to the quality of input data. This approach not only improves the fusion accuracy, but also increases the computational efficiency, leading to the proposal of a new architecture, MIEDCSTFN. 10m-resolution data for the corresponding dates are generated using the output Landsat data from the improved EDCSTFN model as input to the DSTFN model. Comparative validation with several models shows that the improved model has higher accuracy and robustness, and the obtained 10m data are very close to the real data. Compared with the original model, SSIM improves 12.54%, RMSE improves 46.38%, SAM improves 15.46%, ERGAS improves 15.74%, and the experimental results show that the improved model has excellent performance and significant advantages in improving image fusion effect.

DOI:
10.1109/ACCESS.2025.3564968

ISSN: