Land surface temperature (LST) simulated by land surface model (LSM) can maintain spatial integrity and high temporal resolution. However, the relatively low spatial resolution limits the practical applications of LSM-simulated LST. Traditional downscaling methods often require lots of auxiliary data and suffer from significant loss of spatial details under large scale differences. To this end, we propose a super-resolution (SR) downscaling method based on multi-source reference using a multi-attention multi-residual network (MAMRN). The Moderate Resolution Imaging Spectroradiometer (MODIS) LST was used as the reference to improve the spatial resolution of LSM-simulated China Land Data Assimilation System (CLDAS) LST. Six regions with different land cover types covered the Chinese mainland were selected to test the MAMRN’s performance, and a traditional bilinear interpolation method and three deep learning-based SR methods were used for comparison. Comparative experiments demonstrate that MAMRN achieves improved performance, both visually and quantitatively in all six regions. Specifically, the average Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR) and Learned Perceptual Image Patch Similarity (LPIPS) are 2.06 K, 16.76, and 0.54, respectively. The quantitative evaluation scores surpass those of the comparison methods. The transfer experiments and ablation studies also indicate MAMRN’s superiority and effectiveness. Validation by in-situ LST shows that MAMRN can accurately retrieve LST under both clear-sky and cloudy-sky conditions, with the overall accuracy of clear-sky slightly superior to that of cloudy-sky. In a word, MAMRN can enhance the spatial resolution of CLDAS LST (about 6.25 km) to MODIS (1 km) scale while preserving clear spatial details. This capability is beneficial for generating spatially integrity LST with high spatiotemporal resolution, and contributes to the study of global climate change. Our code can be available at https://github.com/AHU-RS/MAMRN.

Shadow removal is beneficial for various remote sensing applications, such as semantic segmentation and object detection. However, traditional shadow removal methods perform poorly when applied to high-resolution remote sensing images. Some deep learning-based models exhibit high-level accuracy of removal, but it is difficult to be expanded to multi-scenarios due to insufficient pairs of shadow/shadow-free data in reality. In this paper, a novel framework is proposed from shadow synthesis to shadow removal, aiming to improve the usability of deep learning-based models in the task of shadow removal in high-resolution remote sensing images. First, we combine the physical shadow illumination model and a domain alignment network to synthesize a large variety of realistic shadows. Then, a shadow removal network considering global-local features is built to restore the ground surface details finely. Numerous experiments have shown that the proposed framework can effectively remove various shadows and is superior to existing methods.

There is no "cloudy & cloud-free" paired images with totally identical surface information under the same spatial and temporal condition in reality, which limits the development of supervised deep learning methods in the field of cloud removal and detection. In this regard, a high-fidelity thin cloud synthetic method is proposed, which is more challenging than the synthesis of thick clouds. This method combines physical model and data-driven methods. The expression of scattering laws at the pixel level is extended to the channel level to synthesize multi-channel cloud from cirrus band with controlled cloud thickness. Besides, spatial and spectral features are learned from real data using generative adversarial networks and transformed to synthetic data. Based on it, a dataset containing infinite number of "cloudy & cloud-free" pairs can be constructed. Experimental results show that the proposed method has the best visual effect with highest quantitative evaluations compared with current methods.

Local climate zone (LCZ) mapping can explore the variability of the impact of urban form on the thermal environment in different urban contexts, and large-scale LCZ mapping can help us to better understand the spatial and temporal dynamics of the climate in urban areas around the world. Studies have indicated that deep learning-based methods can effectively perform LCZ classification. However, the accuracy of LCZ classification on large-scale datasets is still unsatisfactory, mainly due to the fact that the traditional convolutional neural networks are not good at mining contextual information, which is crucial for fully understanding remote sensing scenes. In this paper, to solve this problem, we propose an LCZ mapping method based on remote sensing images by coupling multi-level features mined from global and local ranges with prior knowledge, named LCZ-MFKNet. The global and local features are extracted through Swin Transformer and space-maintained ResNet (SM-ResNet) model branches, respectively, and then fused through an improved squeeze-and-excitation (iSE) module. The prior knowledge studied from the theoretical definition and experimental tests is that two typical sets of LCZ categories are easily confounded in multi-class classification but separable in two-class classification. Experiments are conducted on the large publicly available So2Sat LCZ42 dataset, where the proposed LCZ-MFKNet method achieved the highest LCZ mapping accuracy. Moreover, six megacities were selected globally for LCZ mapping, and the results verified the accuracy and the general applicability of the proposed LCZ-MFKNet method in large-scale LCZ mapping.

At the end of 2019, the very first COVID-19 coronavirus infection was reported and then it spread across the world just like wildfires. From late January to March 2020, most cities and villages in China were locked down, and consequently, human activities decreased dramatically. This letter presents an “offline learning and online inference” approach to explore the variation of PM2.5 pollution during this period. In the experiments, a deep regression model was trained to establish the complex relationship between remote sensing data andin situPM2.5 observations, and then the spatially continuous monthly PM2.5 distribution map was simulated using the Google Earth Engine platform. The results reveal that the COVID-19 lockdown truly decreased the PM2.5 pollution with certain hysteresis and the fine particle pollution begins to increase when advancing resumption of work and production gradually.

Due to the tradeoff between spatial and temporal resolutions of remote sensing images, spatiotemporal fusion models were proposed to synthesize the high spatiotemporal image series. Currently, spatiotemporal fusion models usually employ one coarse-resolution image acquired on a prediction date and at least another pair of coarse–fine resolution images close to the prediction time as references to derive the fine-resolution image on the prediction date. After years of development, the model accuracy has gained a certain improvement, but nearly, all the models require at least three image inputs and rigid time constraints must be applied to the references to guarantee the fusion accuracy. However, it is not always that easy to collect adequate data pairs for fine-resolution image series simulation in practice because of the bad weather condition or the time inconsistency between the coarse–fine resolution data sources, which causes some difficulties in the actual application. This article introduces the conditional generative adversarial network (CGAN) and switchable normalization technique into the spatiotemporal fusion problem and proposes a flexible deep network named the GAN-based SpatioTemporal Fusion Model (GAN-STFM) to reduce the number of model inputs and broke the time restriction on reference image selection. The GAN-STFM just needs a coarse-resolution image on the prediction date and another fine-resolution reference image at an arbitrary time in the same area for model inputs. As far as we know, this is the first spatiotemporal fusion model that requires only two images as model inputs and puts no restriction on the acquisition time of references. Even so, the GAN-STFM performs on par or better than other classical fusion models in the experiments. With this improvement, the data preparation for spatiotemporal fusion tends to be much easier than before, showing a promising perspective for practical applications.

Significant progress has been made in spatiotemporal fusion for remote sensing images; however, most models require inputs to be free of clouds and without missing data, considerably confining their applications in practice. Due to recent advances in deep learning technologies, powerful modeling capabilities could be leveraged to bring potential solutions to this problem. This article proposes a novel architecture named the robust spatiotemporal fusion network (RSFN) based on the generative adversarial network and attention mechanism with dual temporal references to automatically handle input noise. The RSFN only needs one coarse-resolution image on the prediction date and two referential fine-resolution images before and after the prediction date as model inputs. Most notably, there is no special restriction attached on the data quality of referential images. The comparison with other models demonstrates the effectiveness of the RSFN model quantitatively and visually in four study areas using MODIS and Landsat images. Two main conclusions can draw from the experiments. First, the input data noise hardly affects the prediction results of the RSFN, and the RSFN can gain a comparable or even higher accuracy; conversely, the other methods only show limited resistance to input noise. Second, the RSFN with cloud-contaminated references outperforms the other models with cloud-free references after data filtering in the same study area during the same period. The satellite data quality usually varies significantly; the model robustness and fault tolerance are considered critical for actual applications. The RSFN is a simple end-to-end deep model with high accuracy and fault tolerance designed for spatiotemporal fusion with imperfect data inputs, showing promising prospects in practical applications.

Due to the rapid urbanization, the impact of urban spatial form on surface temperature cannot be ignored and is becoming increasingly significant. This study investigates the impact of the urban spatial form which includes building form and urban land surface moisture (ULSM) on the land surface temperature (LST) of communities in Wuhan, China. The LST was retrieved in the summer and winter by Landsat 8 OLI/TIRS data. Thirty typical communities in the three-ring area were selected to represent general residential area in Wuhan. By extracting the building height (BH), building density (BD), floor area ratio (FAR), sky view factor (SVF), frontal area index (FAI) and ULSM of each community, the result shows that BH, BD, FAR, SVF and ULSM are critical variables in lowering LST in both winter and summer.