Due to the tradeoff between spatial and temporal resolutions commonly encountered in remote sensing, no single satellite sensor can provide fine spatial resolution land surface temperature (LST) products with frequent coverage. This situation greatly limits applications that require LST data with fine spatiotemporal resolution. Here, a deep learning-based spatiotemporal temperature fusion network (STTFN) method for the generation of fine spatiotemporal resolution LST products is proposed. In STTFN, a multiscale fusion convolutional neural network is employed to build the complex nonlinear relationship between input and output LSTs. Thus, unlike other LST spatiotemporal fusion approaches, STTFN is able to form the potentially complicated relationships through the use of training data without manually designed mathematical rules making it is more flexible and intelligent than other methods. In addition, two target fine spatial resolution LST images are predicted and then integrated by a spatiotemporal-consistency (STC)-weighting function to take advantage of STC of LST data. A set of analyses using two real LST data sets obtained from Landsat and moderate resolution imaging spectroradiometer (MODIS) were undertaken to evaluate the ability of STTFN to generate fine spatiotemporal resolution LST products. The results show that, compared with three classic fusion methods [the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), the spatiotemporal integrated temperature fusion model (STITFM), and the two-stream convolutional neural network for spatiotemporal image fusion (StfNet)], the proposed network produced the most accurate outputs [average root mean square error (RMSE) < 1.40 °C and average structural similarity (SSIM) > 0.971].

Traditionally, clean reference images are needed to train the networks when applying the deep learning techniques to tackle image denoising tasks. However, this idea is impracticable for the task of synthetic aperture radar (SAR) image despeckling, since no real-world speckle-free SAR data exist. To address this issue, this article presents a noisy reference-based SAR deep learning filter, by using complementary images of the same area at different times as the training references. In the proposed method, to better exploit the information of the images, parameter-sharing convolutional neural networks are employed. Furthermore, to mitigate the training errors caused by the land-cover changes between different times, the similarity of each pixel pair between the different images is utilized to optimize the training process. The outstanding despeckling performance of the proposed method was confirmed by the experiments conducted on several multitemporal data sets, when compared with some of the state-of-the-art SAR despeckling techniques. In addition, the proposed method shows a pleasing generalization ability on single-temporal data sets, even though the networks are trained using finite input-reference image pairs at a different imaging area.

Geostationary satellite land surface temperature (GLST) data are important for various dynamic environmental and natural resource applications for terrestrial ecosystems. Due to clouds, shadows, and other atmospheric conditions, the derived LSTs are often missing a large number of values. Reconstructing the missing values is essential for improving the usability of the geostationary satellite LST data. However, current reconstruction methods mainly aim to fill the values of a small number of invalid pixels with many valid pixels, which can provide useful land surface temperature values. When the missing data extent becomes large, the reconstruction effect will worsen because the relationship between different spatiotemporal geostationary satellite LSTs is complex and highly nonlinear. Inspired by the superiority of the deep convolutional neural network (CNN) in solving highly nonlinear and dynamic problems, a multiscale feature connection CNN model is proposed to fill missing LSTs with large missing regions. The proposed method has been tested on both FengYun-2G and Meteosat Second Generation-Spinning Enhanced Visible and InfraRed Imager geostationary satellite LST datasets. The results of simulated and actual experiments show that the proposed method is accurate to within about 1 °C, with 70% missing data rates. This is feasible and effective for large regions of LST reconstruction tasks.

Super-resolution mapping (SRM) is an effective way to alleviate the mixed pixel problem of remotely sensed imagery, by transforming the coarse-resolution fraction image originated from spectral unmixing into a fine-resolution land-cover map. Deep learning has been widely used in SRM since it has a powerful ability to represent the complex heterogeneous spatial distribution patterns of land-cover patches. However, the accuracy of existing deep learning-based SRM models is compromised by the fact that the fraction images used in SRM always contain errors. In this paper, we propose an end-to-end convolutional neural network (CNN)-based fraction error eliminating SRM (DeepNESRM) method to overcome the negative effect of fraction errors in SRM. In DeepNESRM, to better learn the complex nonlinear relationship between the actual coarse-resolution fraction image and the fine-resolution land-cover map by the CNN, a practical error simulation method that considers the characteristics of fraction errors is introduced to produce training samples. In addition, a multi-level feature fusion CNN is adopted to eliminate fraction errors and simultaneously implement SRM. Experiments using Sentinel-2 and Landsat 8 images were conducted to test the performance of the proposed method. Two conventional SRM methods, namely the pixel swapping method and the spatial dependence and L2 norm combined SRM (L2_SRM) method, and also a stacked very deep CNN based SRM (VDSRM) method, were used as the comparison methods. The results show that DeepNESRM can deal with fraction errors and preserve the spatial detail information, and achieve higher average overall accuracies than the other methods.

Rivers are important elements of the earth’s ecosystem, and their spatial distribution information is critical for the study of hydrological and biogeochemical processes. Moderate Resolution Imaging Spectroradiometer (MODIS) imagery has been widely used for river mapping due to its high temporal resolution and long-term observation records, which are essential to capture the rapid fluctuation of rivers. However, when the conventional hard classification methods are used, the accuracy of the river maps produced from MODIS data (and especially for those rivers with narrow widths) is often limited because mixed pixels are common in MODIS imagery, due to coarse spatial resolution. In this article, a cascaded spectral–spatial information combined deep convolutional neural network (CNN) model for super-resolution river mapping (DeepRivSRM) is proposed to produce Landsat-like fine-resolution river maps from MODIS images. In DeepRivSRM, a CNN made up of a spectral unmixing module and a super-resolution mapping (SRM) module is introduced to handle the spectral and spatial information simultaneously. Moreover, for training the DeepRivSRM model, an adaptive cross-entropy loss function incorporating the fraction information of the rivers is designed to improve the performance of the DeepRivSRM model for small rivers. The proposed method was evaluated with MODIS images from three test sites and was compared with hard classification, a conventional SRM method, and a CNN-based SRM method. The results show that DeepRivSRM can generate more accurate river maps by effectively learning the subpixel-scale spatial–spectral information in MODIS imagery.