Federated learning (FL) has recently been adopted to train shared models across industrial Internet of Things (IoT) devices without revealing their private raw data. Conventional FL usually relies on a central server for coordination. However, in reality, the central server is not fully trusted, which means that it may be collecting data, raising concerns about data leakage and misuse. Here, we propose a decentralized FL algorithm based on deep neural networks to address the problem of untrusted central servers. The original problem is decomposed into several subproblems with consensus constraints, which can be solved by local computation and communication. The proposed algorithm combines the decentralized average consensus and alternating direction method of multipliers. Several decentralized algorithms are employed for comparison, and the issue of heterogeneous data is discussed. Experimental evaluation shows the effectiveness of our proposed algorithm.

The Global Navigation Satellite System reflectometry (GNSS-R) technique has demonstrated its potential for terrestrial applications. Over inland water bodies, the dominance of coherent components in GNSS-R has been widely recognized. Nevertheless, little attention is given to GNSS-R speckle, which is inherent to coherent imaging systems. In this study, taking the multiplicative speckle into account, we regard GNSS-R coherent scattering as a statistical distribution. First, the expression of the statistical distribution is identified and parameterized using observations from the Cyclone Global Navigation Satellite System (CyGNSS). The results suggest that the power tends to obey a three-degrees-of-freedom distribution model. Second, the multilook statistics of CyGNSS, such as the mean value and the coefficient of variation (CV), are analyzed on different spatial–temporal scales. Finally, we realize surface water mapping using multilook statistics. Comparison with the state-of-the-art algorithms shows that the proposed method can effectively improve the goodness of water mapping, with higher overall accuracies (~0.97) and F1 scores (~0.60). This study provides new insights into future GNSS-R land observations.

Gridded precipitation products with fine resolutions and qualities are of great importance for understanding the global water–carbon-energy cycles at various spatiotemporal scales. Though continuous developments in Satellite Remote Sensing fields have been providing great strengths for measuring the precipitation from space, merging precipitation products from different sources, especially the gauge observations, is still the optimal way for obtaining high-quality precipitation data. Currently, the mainstream merging methods mainly focus on merging the rain rates without the considerations of rain events. In this study, we propose a new assumption that both rain events and rain rates should be considered in the merging procedures rather than only the rain rates. To meet our assumption, a morphology-based adaptive spatio-temporal merging algorithm (MASTMA) for combining various precipitation products is proposed, in which the morphology theory is first introduced to comprehensively consider the influences from both rain events and rain rates. The multisource and multiscale precipitation products including the gauge-based data (CPC-U, 0.5°, daily), the satellite-based data [Global Satellite Mapping of Precipitation by Moving Vector with Kalman (GSMaP-MVK), 0.1°, hourly; integrated multisatellite retrievals for global precipitation measurement late run (IMERG-LR), 0.1°, half-hourly], and the reanalysis data (ERA5-land, 0.1°, hourly), have been comprehensively considered in MASTMA for generating the final estimates (MASTMA-F, 0.1°, hourly) over the southeastern regions of the Mainland China in the periods from 2016 to 2019. The main conclusions include but are not limited to: 1) considerations on rain events contribute significantly to the final merged results, especially when eliminating false extreme values over the regions where precipitation is greatly overestimated; 2) the MASTMA could optimally integrate the advantages from multisource precipitation products with different resolutions, particularly from the perspective of the spatial distributions; and 3) the final merged estimates using MASTMA outperform the contemporary state-of-the-art precipitation products especially in terms of modified Kling–Gupta Efficiency (mKGE) and critical success index (CSI). Additionally, the results of this study suggest that MASMTA is a new promising merging approach with great robustness and applicability, and has the foreseeable potentials for the operational run to generate the optimal global merged precipitation products.

Triple collocation (TC) is a novel method for quantifying the uncertainties of three data sets with mutually independent errors and has been widely used over different geographical fields. Researches in recent years report that TC shows potential in merging multiple data sets from different sources, while the TC-based merging method has not been used over precipitation. Using the TC formulation, this study merges precipitation from the Climate Prediction Center's morphing technique (CMORPH), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN), and the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA5). The interim ECMWF Re-Analysis (ERA-Interim) is also involved to act as the substitute of ERA5 in some specific experiments for quality comparison between them. Merged data sets are produced at 0.25°×0.25°and daily resolutions from March 2000 to December 2013 over Mainland China, using ground observations from more than 2000 rain gauges as the validation benchmark. First, the effectiveness of the TC-based method for precipitation merging is assessed. Then, two weighting methods using root-mean-square error (RMSE) in logarithmic scale (log-RMSE) and modified scale (mod-RMSE) are compared because previous studies show that mod-RMSE is more suitable for characterizing errors within estimated data. Meanwhile, two merging strategies are designed, that is, merging rainfall and snowfall separately (RS) and merging precipitation directly (P). The results show that 1) all the merged products are superior to any input product which proves that the TC method is effective in precipitation merging; 2) TC-based merging generally has a better performance than dynamic Bayesian model averaging (DBMA)-based merging; 3) mod-RMSE shows worse performance in weight estimation than log-RMSE because mod-RMSE will deteriorate the impact of the underestimated inputs; and 4) RS-based merging is superior to P-based merging, and the superiority is particularly notable in winter. The RS strategy will be very helpful in improving the accuracy of precipitation estimates in cold climate such as over mountainous and high-altitude regions. Finally, the limitations of the TC method and potential solutions are discussed. This study demonstrates the great potential of the TC-based merging method in precipitation and provides insights into its application and development.

Precipitation plays a critical role in the water and energy cycle. The systematic and random errors of precipitation are usually estimated using the additive model. However, various studies have shown that the multiplicative model is more suitable to describe the errors of precipitation than the additive model. This study integrates the multiplicative model with the Willmott-AghaKouchak method to characterize the errors of four selected representative satellite precipitation products in China. Zero precipitation is addressed by adding a tiny increment, which is determined by a sensitivity analysis, enabling the examination of missed precipitation and false alarms compared with the traditional strategy that only considers hit events. The results show that the systematic errors based on the additive model are too sensitive to heavy precipitation, resulting in problems, such as unexpected fluctuations, regional biases, unsteady performance, and reverse seasonal and elevational trends in some cases. In contrast, the multiplicative model resolves these problems through balancing the contributions of light and heavy precipitation and is recommended for systematic and random error estimation.

Infrared (IR) information is fundamental to global precipitation estimation. Although researchers have developed numerous IR-based retrieval algorithms, there is still plenty of scope for promoting their accuracy. This article develops a novel deep learning-based algorithm entitled infrared precipitation estimation using a convolutional neural network (IPEC). Based on the five-channel IR data, the IPEC first identifies the precipitation occurrence and then estimates the precipitation rates at hourly and 0.04° × 0.04° resolutions. The performance of the IPEC is validated using the Stage-IV radar-gauge-combined data and compared to the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Cloud Classification System (PERSIANN-CCS) in three subregions over the continental United States (CONUS). The results show that the five-channel input is more efficient in precipitation estimation than the commonly used one-channel input. The IPEC estimates based on the five-channel input show better statistical performance than the PERSIANN-CCS with 34.9% gain in Pearson's correlation coefficient (CC), 38.0% gain in relative bias (BIAS), and 45.2% gain in mean squared error (MSE) during the testing period from June to August 2014 over the central CONUS. Furthermore, the optimized IPEC model is applied in totally independent periods and regions, and still achieves significantly better performance than the PERSIANN-CCS, indicating that the IPEC has a stronger generalization capability. On the whole, this article proves the effectiveness of the convolutional neural network (CNN) combined with the physical multichannel inputs in IR precipitation retrieval. This end-to-end deep learning algorithm shows the potential for serving as an operational technique that can be applied globally and provides a new perspective for the future development of satellite precipitation retrievals.

Robust validation of the ensemble precipitation datasets was vital for assessing and then improving the quality and confident application in water energy cycle related research. In this study, the performance of the Ensemble Multi-Satellite Precipitation Datasets using the Dynamic Bayesian Model Averaging scheme (EMSPD-DBMA) was evaluated and compared against the rain gauge networks, as well as the Day-1 Level 3 products of the newly launched of Global Precipitation Measurement (GPM) mission over the Tibetan Plateau (TP). Comprehensive analyses of daily precipitation estimates for 2000-2015 showed that the EMSPD-DBMA products showed a reliable performance in estimating the precipitation regimes in the past 16 years over the TP, where the averaged correlation coefficients and relative bias (RB) were 0.532 and -8.9%, respectively. The EMSPD-DBMA datasets appreciably showed better correlations and lower errors than that of GPM in the summer season of 2014 and 2015, though with very similar spatial patterns. Moreover, the ensemble products EMSPD-DBMA could significantly improve the rainfall detection around 1.3 times GPM-era products at the survey periods.

In the summer of 2016, severe storms caused serious casualties and destruction of facilities and properties over South China. Near-real-time (NRT) satellite precipitation products are attractive to rainstorm monitoring and flood warning guidance owing to its combination of timeliness, high spatiotemporal resolution, and broad coverage. We evaluate the performance of four NRT satellite products, i.e., Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks, 3B42RT, Global Satellite Mapping of Precipitation (GSMaP) NRT, and Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG) Late run using a high-quality merged product in the rainy June over South China. In addition, a method based on an empirical flash flood guidance and the Flash Flood Potential Index is proposed to examine the applicability of satellite products in guiding flood warning. The IMERG Late run and GSMaP NRT perform the closest-to-ground observations. 3B42RT detects the most flood warning events due to its notable overestimation of actual precipitation. We recommend that the IMERG Late run is the best NRT satellite product in capturing flood hazard events according to the Pareto Efficiency of jointly optimizing higher hit ratio and lower false alarms.

Two satellite precipitation products (3B42RT and 3B42V7) of the Tropical Rainfall Measuring Mission (TRMM) show systematic overestimation anomalies over inland water bodies in high mountain Asia (HMA). The relative difference (RD) was calculated between precipitation estimates over pixels containing water bodies and over pixels neighboring water bodies. A t-test was employed to check the statistical significance. Results show that ~53% of the water bodies passed the significance test at the daily scale for the TRMM products on average. Gauge adjustment alleviates the overestimation of 3B42V7 (mean RD = 20%) compared with 3B42RT (mean RD = 34%). Liquid water surfaces tend to affect the quality of TRMM Multisatellite Precipitation Analysis more significantly than ice water surfaces due to different passive microwave (MW) emission characteristics. In contrast, Global Precipitation Measurement (GPM) mission products provide more consistent precipitation estimates, with only ~4% of the water bodies passing the significance test. TRMM-based precipitation estimates are much higher than GPM products over inland water bodies, which is, however, not the case over the land. The improvement is mainly attributed to the unified and updated MW algorithm used in GPM products.