Accurate flood forecasting remains a critical challenge for water-resource management, as it demands simultaneous modeling of local, time-varying runoff drivers (e.g., rainfall-induced peaks, base- flow trends) and complex spatial interactions across a river network. Traditional data-driven approaches, such as convolutional networks and sequence-based models, ignore topological information about the region. Graph Neural Networks (GNNs), in contrast, propagate information exactly along the river network, making them ideal for learning hydrological routing. However, state-of-the-art GNN-based flood prediction models still collapse pixels to coarse catchment polygons because the cost of training explodes with graph size and higher resolution. Furthermore, most existing methods treat spatial and temporal dependencies separately, either applying GNNs solely on spatial graphs or transformers purely on temporal sequences, thus failing to simultaneously capture spatiotemporal interactions critical for accurate flood prediction. To address these limitations, we introduce a heterogenous basin graph to represent every land and river pixel as a node connected by both physical hydrological flow directions as well as inter-catchment relationships. We also propose HydroGAT, a novel spatiotemporal network that adaptively learns both local temporal importance as well as most influential upstream locations. Evaluated in two Midwestern US basins and across five baseline architectures, our model achieves higher NSE (up to 0.97), improved KGE (up to 0.96), and low bias (PBIAS within ± 5%) in hourly discharge prediction, while offering interpretable attention maps that reveal sparse, structured intercatchment influences. To support high-resolution basin-scale training, we develop a distributed data-parallel pipeline that scales efficiently up to 64 NVIDIA A100 GPUs on NERSC Perlmutter supercomputer, demonstrating up to 15× speedup across machines. Our code is available at https://github.com/swapp-lab/HydroGAT.

Data assimilation has been demonstrated as the potential crop yield estimation approach. Accurate quantification of model and observation errors is the key to determining the success of a data assimilation system. However, the crop growth model error is not fully taken into account in most of the previous studies. The objective of this study is to better quantify the model uncertainty in the data assimilation system. Firstly, we calibrated a crop growth model and inferred its posterior uncertainty based on the Global LAnd Surface Satellite (GLASS) 250-m LAI product, regional statistical data, station observations, and field measurements with a Markov chain Monte Carlo (MCMC) method. Secondly, the model posterior uncertainty was used in the Ensemble Kalman Filter (EnKF) algorithm to better characterize the ensemble distribution of model errors. Our results indicated the proposed Bayesian posterior-based EnKF can improve the accuracy of winter wheat yield estimation at both the point scale (the coefficient of determination R2value increasing from 0.06 to 0.41, the mean absolute percentage error MAPE value decreasing from 12.65% to 7.82%, and the root mean square error RMSE value decreasing from 987 to 688 kg∙ha-1) and the regional scale (R2value from 0.30 to 0.57, MAPE value from 19.67% to 10.13%, and RMSE value from 1275 to 695 kg∙ha-1) compared with the open-loop estimation. Our analysis also indicated that the Bayesian posterior-based EnKF can perform better compared to the standard Gaussian perturbation-based EnKF. The proposed framework provides an important reference for crop yield estimation at the regional scale in similar agricultural landscapes worldwide.

Against rapid development in urban areas, timely urban land cover changes (ULCC) information is beneficial for understanding the urban environment and promoting sustainable development. To realize real-time urban land cover change detection, high-frequency remotely sensed data are urgently needed. In this study, we tested the detection capability of urban land cover changes using a new daily nighttime light image (Black Marble). Firstly, time series of VNP46V2 from 2012-2019 were collected and decoded into annual trend segments using the BFAST Monitor model. Then, we recognized the jump point in trend segments and defined the corresponding pixel as urban land cover change. We analyzed the Normalized Difference Vegetation Index (NDVI) time series from Landsat images spanning 2014-2019 and removed pixels without significant seasonal fluctuations from ULCC assembled. Finally, the magnitude and change time of ULCC pixels were quantified through BFAST Monitor. It was proved that Black Marble performed well in ULCC detection, achieving an overall accuracy of 87.75%, and detected change time was accurate to 81.12% under ± 1 year allowable error. The present Black Marble data have the potential for real-time urban land use detection and global mapping of ULCC, especially in areas without enough clear-sky observations.

In this work, we compare the impact of two ways of representing the crop growth model uncertainty (the crop growth model was calibrated by Markov Chain Monte Carlo (MCMC)) on the subsequent assimilation of the Sentinel-2 LAI through Ensemble Kalman Filter (EnKF): generating N sets of crop parameters based on the Optimal-posterior parameter value (reference method, we name it OMCMC-EnKF) and Randomly selecting N sets of crop parameters from the most frequent parameters (our proposed method, RMCMC-EnKF). The results show that our proposed method is better than the reference model, when compared with the field measured winter wheat LAI and yield, its performance in the R2, RMSE, and uncertainty estimation is better. Furthermore, our proposed method is more robust to feedback from remote sensing data and more resistant to lower quality observations encountered during assimilation.

Accurate and timely regional crop yield information, particularly field-level yield estimation, is essential for commodity traders and producers in planning production, growing, harvesting, and other interconnected marketing activities. In this study, we propose a novel data assimilation framework. Firstly, we construct the county-level prior and likelihood constraints for a process-based crop growth model based on the previous year's statistical yield and the current year's field observations. Then, we infer the posterior sets of model-simulated time-series LAI and the final yield of winter wheat with an MCMC (Markov chain Monte Carlo) method for each meteorological data grid of ERA5 (European Centre for Medium-Range Weather Forecasts Reanalysis v5). Finally, we estimate the winter wheat yield at the spatial resolution of 10 m by combining Sentinel-2 LAI and the WOFOST model in Hengshui, the prefecture-level city of Hebei province of China. The results show that the proposed framework can estimate the winter wheat yield with a coefficient of determination R2equal to 0.29 and mean absolute percentage error MAPE equal to 7.20% compared with field measurements. However, agricultural stress that crop growth models cannot quantitatively simulate, such as lodging, can greatly reduce the accuracy. The results also suggest good agreements with county-level statistics of the growing year with a coefficient of determination R2equal to 0.52 and mean absolute percentage error MAPE equal to 7.19%.

Assessing crop lodging at the regional scale is an important requirement for breeding lodging-resistant varieties and harvest planning. Accurately and continuously estimating crop lodging area from remote sensing data remains challenging due to the high randomness scattering signal of SAR images and the insufficient number of applicable optical images. This study developed a new framework for estimating crop lodging area based on SAR data using the spatial aggregation approach of field units, overcoming the deficit of the traditional pixel-based approach susceptible to speckle noise and spatial heterogeneity. We aggregated the field’s pixel in SAR images using the spatial aggregation approach. The lodging area estimation models of dual-pol and quad-pol were established using a random forest (RF) algorithm. The Sobol approach evaluated the uncertainty and sensitivity at a regional scale. Finally, we analyzed the scattering mechanisms of the lodging field. Results indicate that the proposed method achieves the high performance of the lodging area estimates at the regional scale, in the testing set, with R2and RMSE of the GF-3 model being 0.57 and 18.63%, and the Sentinel-1 model is 0.49 and 20.59%. Besides, the uncertainties of models are below 10%, and are insensitive to the variation of parameters inter-correlation. The depolarization effect and scattering randomness gradually weaken with the increase of lodging percentages. In contrast, the surface scattering quickly increases and finally dominates the total scattering after lodging percentages greater than 80%. This proposed approach would help develop a real-time crop lodging monitoring system using SAR data.

Spatially distribution information on wheat yield forecasting at the large regional scale is important for market forecast and agricultural sustainable development. Assimilating remote sensing information into the crop growth model has demonstrative to be the effective approach for crop yield prediction. However, it remains a challenge to determine the crop growth model of input parameters and initial conditions at the spatial regional scale. In the paper, we proposed a Markov Chain Monte Carlo (MCMC) and 4DVAR hierarchical data assimilation scheme, which enables the winter wheat yield forecasting at the 500 m grid size ahead of harvest time in Henan province. This approach applies data assimilation algorithms at two spatial scales. At the county scale, the MCMC algorithm was used to recalibrate the uncertain and sensitive parameters of the WOFOST model using the county-level statistical yield. Then, we assimilated time-series MODIS reflectance into WOFOST-PROSAIL coupled model using the 4DVAR method for each 500 m pixel across the entire Henan province of China. The results show that the simulated yield was strong positive correlated with the statistical yield at county-level scale with R2 = 0.81 and RMSE = 877 kg/hm2, which demonstrated the potential usage of the MCMC-4DVAR based large area yield estimation with remote sensed data and yield statistics.

Crop yield estimation at the field scale is essential for farmers, crop insurance companies to make informed decisions. Methodologies based on assimilating remote sensing LAI into crop growth models have shown advantages in crop yield estimates. Compared with MODIS and Landsat, Sentinel-2 satellites provide higher spatial and temporal resolution data, which brings revolutionary opportunities for crop monitoring. This study is to evaluate the performance of assimilating Sentinel-2 LAI into the WOFOST model for winter wheat yield estimation using the Ensemble Kalman Filter algorithm. The results showed that assimilating Sentinel-2 LAI improved the yield estimation (R2= 0.45; RMSE = 512 kg/ha) compared to the situation without data assimilation (R2= 0.27; RMSE = 818 kg/ha), which demonstrated the potential usage of the Sentinel-2 LAI for yield estimation at the field scale.