Fusing Landsat-7, Landsat-8 and Sentinel-2 surface reflectance to generate dense time series images with 10m spatial resolution
- 4 March 2022
- journal article
- research article
- Published by Taylor & Francis in International Journal of Remote Sensing
- Vol. 43 (5) , 1630-1654
Abstract
Due to the influence of the external atmospheric environment and the physical constraints of the sensors, it is difficult for existing medium spatial resolution sensors to produce intensive time-series images. The easy accessibility and similar sensor configuration of Landsat and Sentinel-2 data make them suitable for synthesizing time-series images. Currently, some deep learning-based studies have achieved significant success in multi-sensor fusion, but these deep-learning fusion networks only use single features that are not enough to produce intensive time-series images. Shallow features are easy to extract but have weak representation, while the deep network can extract abstract features with strong representation ability but faces the problem of gradient disappearance and model degradation. To address these problems, we propose a novel spatial-temporal-spectral integration model that can simultaneously fuse Landsat-7,8 and Sentinel-2 sensors. First, an enhanced residual dense network (ERDN) is proposed to improve Landsat’s four bands (Blue, Green, Red, and Near-infrared) to 10 m spatial resolution. ERDN’s residual dense structure utilizes abundant shallow features (i.e. texture and spectral information) and deep features (i.e. angles of view and aspect ratios of objects) to reconstruct spatial structures. At the same time, the introduced high-resolution images can provide reliable spatial auxiliary information for the missing image parts. Second, we employ a time-series-based reflectance adjustment (TRA) method to reduce the reflectance differences between Landsat and Sentinel-2 images. The fusion results indicate that our fusion framework can not only fill the gap caused by the Landsat-7 scan-line corrector (SLC) failure but also restore more spatial details than previous methods. More importantly, it can predict tiny reflectance changes in land cover. Overall, our study can produce continuous reflection observations with a higher time density and provides important data guarantee for extracting dynamic changes in land coverages.Keywords
Funding Information
- The National Natural Science Foundation of China (No. 42001327)
- the China Postdoctoral Science Foundation (No. 2019M660003)
This publication has 31 references indexed in Scilit:
- A flexible spatiotemporal method for fusing satellite images with different resolutionsRemote Sensing of Environment, 2016
- Estimation of rice phenology date using integrated HJ-1 CCD and Landsat-8 OLI vegetation indices time-series imagesJournal of Zhejiang University-SCIENCE B, 2015
- Image Super-Resolution Using Deep Convolutional NetworksIEEE Transactions on Pattern Analysis and Machine Intelligence, 2015
- Analysis of sub-pixel snow and ice extent over the extratropical Andes using spectral unmixing of historical Landsat imageryRemote Sensing of Environment, 2014
- Comparison of Commonly Used Image Interpolation Methods2013
- An enhanced spatial and temporal adaptive reflectance fusion model for complex heterogeneous regionsRemote Sensing of Environment, 2010
- A new data fusion model for high spatial- and temporal-resolution mapping of forest disturbance based on Landsat and MODISRemote Sensing of Environment, 2009
- On the blending of the Landsat and MODIS surface reflectance: predicting daily Landsat surface reflectanceIEEE Transactions on Geoscience and Remote Sensing, 2006
- Landsat sensor performance: history and current statusIEEE Transactions on Geoscience and Remote Sensing, 2004
- Image quality assessment: from error visibility to structural similarityIEEE Transactions on Image Processing, 2004