To address the communication challenges associated with Federated Learning (FL), Decentralized Federated Learning (DFL) eliminates the central server and trains the model with decentralized method, enabling each client to only communicate with its neighbors. However, per our analysis, model trained with DFL experiences performance degradation because of data-heterogeneity and time-varying topologies. To address these issues, we propose a Dynamic K-step Gradient Tracking (DKGT) method to enhance the performance of DFL over time varying networks. Specifically, DKGT employs K-step local updates and gradient tracking to reduce the communication cost and the variance from heterogeneous data distribution, and we use dynamic gradient tracking parameter to correct gradient over time varying graph. Theoretically, we derive a universal convergence rate for smooth and non-convex problem at the rate of$\mathcal{O}\left(\frac{\left(f(\textbf{x}_0)-f(\textbf{x}^*)\right)}{\sqrt{T}(L\sqrt{KN})^{-1}-\tau(pKL\sqrt{TKN})^{-1}}+\frac{\sigma^2}{KTN\tau(pK-\tau)}\right)$, that τ and p respectively represent the time window length and the connectivity of time-varying networks. Experimentally, we illustrate the robustness and effectiveness of this heterogeneity correction on extensive non-convex neural network training tasks over different topologies and dynamic network settings.

The fuzzy broad learning system (FBLS) has received increasing attention due to its ability to quickly train from broad learning systems (BLS) and interpretability with fuzzy inference. However, the randomness of BLS brings instability to the training performance of the model, so the hyperparameters of the model are crucial for its performance. Currently, many FBLS use grid search to determine hyperparameters. However, grid search brings longer search time and the parameters obtained have randomness, which may not necessarily be the optimal hyperparameters. In response to these challenges, this paper proposes a fuzzy broad learning system with automatic determination of hyperparameters (ADHFBLS). We construct a novel FBLS based on the interval type-2 fuzzy set and design an incremental learning algorithm for rules and enhancement nodes to support rapid model expansion. Meanwhile, a heuristic hyperparameter automatic optimization algorithm is designed to overcome the randomness and long optimization time of grid search. Experiments have shown that ADHFBLS has higher accuracy and shorter model tuning time compared to some state-of-the-art models based on FBLS.

Subpixel mapping (SPM) is a prevailing technique for addressing the mixed pixel issue by estimating the subpixel-scale distribution of land cover types within mixed pixels. Recently, spatiotemporal subpixel mapping (STSPM) has advanced traditional SPM by integrating prior fine-resolution (FR) images to alleviate the uncertainty in SPM results. However, most existing STSPM methods are typically model-driven or meticulously handcrafted and they usually struggle to flexibly characterize the diverse and complex patterns of land cover patches. This article proposes a novel deep spatiotemporal subpixel mapping network (STSPMNet) that integrates a prior FR land cover map with bitemporal coarse-resolution (CR) remote sensing images using deep learning and change detection techniques. STSPMNet first combines super-resolution and semantic segmentation to achieve the end-to-end SPM from CR remote sensing images to FR land cover maps. A change-detection-based optimization approach is then developed to incorporate a prior FR land cover map with the estimated FR land cover probabilities to generate the final FR land cover map at the prediction date. Three experiments were conducted to evaluate the effectiveness of the proposed STSPMNet. Experimental results demonstrate that STSPMNet outperforms two SPM methods and two STSPM method by recovering more spatial details of land cover patches and generating higher accuracy metrics. Hence, STSPMNet offers a robust solution for producing FR land cover maps from CR remote sensing imagery.

Spatiotemporal remote sensing image fusion (STF) is a promising way to obtain remote sensing data with both fine spatial and temporal resolutions. Gradual and abrupt changes in land surface reflectance images are the main challenges in existing STF methods. Advanced deep learning techniques present powerful ability in learning image-changed information. Therefore, this article proposes a novel spatiotemporal image fusion method using multiscale two-stream convolutional neural networks (STFMCNNs). Multiscale two-stream convolutional neural networks are proposed to capture different sizes of objects in feature learning from a coarse spatial resolution image and two pairs of coarse and fine spatial resolution (FR) images at other dates. Meanwhile, temporal dependence and temporal consistency are explored as complementary information for STFMCNN. Moreover, a local fusion method is developed to characterize local variation by combining two predicted images derived from each stream. Two experiments on different real images are conducted to demonstrate the effectiveness of STFMCNN. Results show that STFMCNN outperformed three existing methods by predicting more accurate FR images with more preserved changed information.

This paper proposes a new subpixel mapping (SPM) method based on multiscale spatial dependence (MSD). At the beginning, it adopts object-based and pixel-based soft classifications to generate the class proportions within each object and each pixel, respectively. Then, the object-scale spatial dependence of land cover classes is extracted from the class proportions of objects, and the combined spatial dependence at both pixel scale and subpixel scale is obtained from the class proportions of pixels. Furthermore, these spatial dependences are fused as the MSD for each subpixel. Last, a linear optimization model on each object is built to determine where the land cover classes spatially distribute within each mixed object at subpixel scales. Three experiments on two synthetic images and a real remote sensing image are carried out to evaluate the effectiveness of MSD. The experimental results show that MSD performed better than four existing SPM methods by generating less isolated classified pixels than those generated by three pixel-based SPM methods and more land cover local details than that generated by an object-based SPM method. Hence, MSD provides a valuable solution to producing land cover maps at subpixel scales.

Superresolution mapping (SRM) is a widely used technique to address the mixed pixel problem in pixel-based classification. Advanced object-based classification will face a similar mixed phenomenon-a mixed object that contains different land-cover classes. Currently, most SRM approaches focus on estimating the spatial location of classes within mixed pixels in pixel-based classification. Little if any consideration has been given to predicting where classes spatially distribute within mixed objects. This paper, therefore, proposes a new object-based SRM strategy (OSRM) to deal with mixed objects in object-based classification. First, it uses the deconvolution technique to estimate the semivariograms at target subpixel scale from the class proportions of irregular objects. Then, an area-to-point kriging method is applied to predict the soft class values of subpixels within each object according to the estimated semivariograms and the class proportions of objects. Finally, a linear optimization model at object level is built to determine the optimal class labels of subpixels within each object. Two synthetic images and a real remote sensing image were used to evaluate the performance of OSRM. The experimental results demonstrated that OSRM generated more land-cover details within mixed objects than did the traditional object-based hard classification and performed better than an existing pixel-based SRM method. Hence, OSRM provides a valuable solution to mixed objects in object-based classification.

Soil moisture (SM) plays an important role in the land surface energy balance and water cycle. Microwave remote sensing has been applied widely to estimate SM. However, the application of such data is generally restricted because of their coarse spatial resolution. Downscaling methods have been applied to predict fine-resolution SM from original data with coarse spatial resolution. Commonly, SM is highly spatially variable and, consequently, such local spatial heterogeneity should be considered in a downscaling process. Here, a hybrid geostatistical approach, which integrates geographically weighted regression and area-to-area kriging, is proposed for downscaling microwave SM products. The proposed geographically weighted area-to-area regression kriging (GWATARK) method combines fine-spatial-resolution optical remote sensing data and coarse-spatial-resolution passive microwave remote sensing data, because the combination of both information sources has great potential for mapping fine-spatial-resolution near-surface SM. The GWATARK method was evaluated by producing downscaled SM at 1-km resolution from the 25-km-resolution daily AMSR-2 SM product. Comparison of the downscaled predictions from the GWATARK method and two benchmark methods on three sets of covariates with in situ observations showed that the GWATARK method is more accurate than the two benchmarks. On average, the root-mean-square error value decreased by 20%. The use of additional covariates further increased the accuracy of the downscaled predictions, particularly when using topography-corrected land surface temperature and vegetation-temperature condition index covariates.

Error source analyses are critical for the satellite-retrieved surface net longwave radiation products. In this study, we evaluate the error sources of the Clouds and the Earth's Radiant Energy System (CERES) project Single Scanner Footprint (SSF) net longwave radiation (NLW) product (monthly mean, 1° equal area) at 11 sites from July 2000 to December 2007 in China. Results show that cloud fraction, land surface temperature and atmospheric temperature error dominate the NLW error, with respective error contributions of ~-20, ~15, and ~10 W/m2, while the total precipitable water vapor has a weak influence. Spatially, due to strong algorithm sensitivity and large product errors, surface temperature and cloud fraction have the most contributions to error sources, especially in northern China, while the atmospheric temperature significantly affects NLW in southern China because of its large product errors. To improve NLW, the large errors of these parameters should be reduced. Our study identifies the dominant error sources and should be helpful to improving the CERES NLW in China.

With increasing of the spatial resolution of satellite imaging sensors, object-based image analysis (OBIA) has been gaining prominence in remote sensing applications. However, scale selection in multi-scale segmentation and OBIA remains a challenge, which directly reduces efficiency of land cover mapping. In this study, we presented an object-based land cover mapping using adaptive scale segmentation. Central to our method is the use of inherent features of segmented objects to determine whether an object should be segmented with a small scale in a top-down segmentation procedure. We firstly used inherent features of a segmented object to determine whether this object should be segmented with a smaller scale in a top-down segmentation procedure, producing a segmentation map with optimal scales. Then, an object-based SVM classifier was applied on the adaptive-scale segmentation map to yield a land-cover map. We have applied this method on a ZY-3 multi-spectral satellite image to produce land cover map, compared with the results using the traditional mean shift algorithm with fixed scales. The experimental results illustrate that the proposed method is practically helpful and efficient to improve the performance of land cover mapping.

This paper proposes spatial distribution pattern-based subpixel mapping (SPMS) as a novel subpixel mapping (SPM) strategy. It separately considers spatial distribution patterns of different types of geographical objects. Initially, it classifies geographical objects into areal, linear, and point patterns according to their spatially geometric characteristics. For the different patterns, SPMS uses the vectorial boundary-based SPM algorithm with the spatial dependence assumption to deal with areal objects, the linear template matching-based SPM algorithm for linear objects, and the spatial pattern consistency matching-based SPM algorithm for point objects. The three patterns are integrated to generate a subpixel map. An artificially created image and two remotely sensed images were used to evaluate the performance of SPMS. The results were compared with a traditional hard classifier and seven existing SPM methods. The experimental results demonstrated that SPMS performed better than the hard classification and traditional SPM methods, particularly when dealing with linear and point objects.

A new superresolution mapping (SRM) method based on high-accuracy surface modeling (HASM) is proposed to generate land-cover maps at the subpixel scale. HASM uses the fundamental theorem of surfaces to uniquely define a land surface, which can produce less errors in interpolation results than classic methods, and thus, the proposed SRM method first uses it to estimate the soft class values of subpixels according to the fraction images of soft classification. Then, it transforms the soft class values into a hard-classified land-cover map using class allocation under the constraints of fraction images. Experiments on a synthetic image and a real remote sensing image show that the proposed method produces more accurate SRM maps than four existing SRM methods. Hence, the proposed method provides a new option for superresolution land-cover mapping.