Hourly monitoring of ground-level fine particulate matter (PM 2.5 ) concentrations forms the basis to assess the short-term PM 2.5 exposure and make rapid responses to pollution events. Satellite remote sensing and ground monitoring stations are able to measure hourly PM 2.5 concentrations, but both of them have strengths and weaknesses: the former features wide spatial coverage, whereas displaying a discontinuous timeline as the retrievals have numerous gaps; conversely, the latter allows for temporally continuous monitoring, but with a limited spatial range around stations being reflected. Thus, efforts are required to map ground-level PM 2.5 at an hourly scale with spatiotemporal continuity. In this article, we developed a framework to generate hourly seamless PM 2.5 estimates by integrating the aforementioned two data sources with complementary spatiotemporal traits. The satellite-derived aerosol optical depth acquisitions are converted along with auxiliary predictors to retrieve ground-level PM 2.5 , and then the missing gaps in the retrievals are filled by fusing the satellite-based retrievals and station-based measurements. Meanwhile, we proposed a promising approach to fill the gaps by combining an adapted spatiotemporal fusion model and an error correction method. The validity of the proposed method is confirmed by mapping hourly PM 2.5 distributions for 2016 in the Wuhan urban agglomeration, China. The proposed reconstruction method achieved R 2 (root-mean-square error) of 0.87 (6.50 μg/m 3 ) and 0.82 (15.01 μg/m 3 ) in the area-based and point-based evaluation, respectively, indicating an excellent model performance. The presented framework maps hourly ground-level PM 2.5 with spatiotemporal continuity and satisfactory accuracy, and represents an important step towards near real-time monitoring.
1 October 2020 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Spatiotemporal fusion of multisource observations is widely accepted for generating dense time series with high spatial resolution. Normally, the fusion process requires at least three input images, i.e., the coarse resolution image on target date and a pair of fine and coarse resolution images on reference date. However, this requirement may not be satisfied, as the coarse resolution image on reference date is potentially unavailable in some applications. This calls for efforts to allow fusion with only the other two input images. So far, the fusion performance under different input configurations has not been fully analyzed. This study is aimed at solving this problem. Specifically, a universal method is proposed to adapt the fusion models requiring three input images to work with two images; then, three popular fusion models are adopted to demonstrate the difference of fusion performance under different input configurations. 38 tests were conducted with MODIS, Landsat, and Sentinel-2 land surface reflectance products. Results suggest, if the high-quality coarse resolution image on reference date is available, the case incorporating three input images leads to slight but insignificant precision superiority over the other one; otherwise, if it is unavailable, the case with only two input images gains considerably better performance. This study challenges the stereotype that spatiotemporal fusion strictly requires at least three input images. The proposed method extends the potential applications of spatiotemporal fusion, and create chances for the existing models to fuse sensors with barely overlapping temporal coverages, such as Landsat 8 OLI and Sentinel-2 MSI.
15 June 2023 in IEEE Geoscience and Remote Sensing Letters
Dense medium-resolution images play an important role in time-series geoscience applications. However, due to technical limitations, remote sensing imaging systems inevitably trade off temporal frequency and spatial swaths, resulting in difficulties to acquire images simultaneously with high spatial and temporal resolution. To overcome this limitation, under the framework of residual learning, we propose a Cross-attention based Adaptive weighting Fusion nEtwork (CAFE) for MODIS-Landsat spatiotemporal fusion to generate dense medium-resolution images. Based on the cross-attention mechanism, we propose multi-channel separated cross-attention and full-feature joint cross-attention blocks to enhance spatial resolution and retain spectral signatures from the perspectives of band-wise processing and full-feature joint processing, respectively. The adaptive temporal difference weighting mechanism is proposed to improve the ability to capture dynamic land surface changes. Besides, we employ an adaptive fusion loss function to constrain the network training. Experimental results indicate that the developed method is superior to several existing algorithms in terms of visual evaluation and quantitative evaluation and it can generate high-quality fusion results in scenarios of both subtle and dramatic temporal changes. Codes will be available at https://github.com/LiupengLin/CAFE.
Medium- to high-resolution imagery is indispensable for various applications. Combining images from Landsat 8 and Sentinel-2 can improve the accuracy of observing dynamic changes on the Earth’s surface. Many researchers use Sentinel-2 10 m resolution data in conjunction with Landsat 8 30 m resolution data to generate 10 m resolution data series. However, current fusion techniques have some algorithmic weaknesses, such as simple processing of coarse or fine images, which fail to extract image features to the fullest extent, especially in rapidly changing land cover areas. Facing the aforementioned limitations, we proposed a multiscale and attention mechanism-based residual spatiotemporal fusion network (MARSTFN) that utilizes Sentinel-2 10 m resolution data and Landsat 8 15 m resolution data as auxiliary data to upgrade Landsat 8 30 m resolution data to 10 m resolution. In this network, we utilized multiscale and attention mechanisms to extract features from coarse and fine images separately. Subsequently, the features outputted from all input branches are combined and further feature information is extracted through residual networks and skip connections. Finally, the features obtained from the residual network are merged with the feature information of the coarsely processed images from the multiscale mechanism to generate accurate prediction images. To assess the efficacy of our model, we compared it with existing models on two datasets. Results demonstrated that our fusion model outperformed baseline methods across various evaluation indicators, highlighting its ability to integrate Sentinel-2 and Landsat 8 data to produce 10 m resolution data.