Reliable drought monitoring in gauge-limited regions requires precipitation products that reproduce both monthly rainfall and the distributions used to derive drought indices. This study developed a 0.1 degrees monthly New Hybrid Precipitation Product (NHPP) for Eswatini for 1986-2024 by merging AgERA5, ERA5-Land, ARC2 and CHIRPS at various staged configurations using Random Forest and XGBoost, followed by seasonal quantile mapping. Performance was assessed through fully nested leave-one-station-out validation and three independent test stations, supported by SHAP interpretation, station quantile analysis, SPI evaluation and NDVI response. The selected model configuration achieved KGE = 0.82, RMSE = 38.86 mm, RBIAS = -2.38% and R-2 = 0.70. GPCC showed slightly better mean station-level Q-Q agreement than NHPP (0.15 versus 0.17), while NHPP was the best-performing machine-learning product. Both datasets reproduced the principal drought periods, with SPI correlations of 0.70-0.84. NHPP identified droughts more frequently, whereas GPCC generally produced longer and more severe SPI-6 and SPI-12 events. NHPP also showed higher NDVI correlations for wet-season SPI-6 and annual SPI-12. NHPP therefore provides finer spatial detail and locally responsive drought information, while GPCC remains a stable regional benchmark; their complementary use improves drought monitoring in data-scarce environments.
ABSTRACT Frequent and severe droughts driven by climate change highlight the need for reliable reservoir operation strategies. While most studies evaluate firm yield at the scale of individual reservoirs, few explicitly consider reservoir system operation of multiple reservoirs, despite its potential to enhance downstream supply reliability. The main objectives are to develop a method to estimate the firm yield of the Seomjingang and Juam Reservoirs in South Korea under reservoir system operation and to define and estimate the minimum inflow time series for each reservoir as a basis for assessing water supply reliability during extreme droughts. Results show that reservoir system operation increases firm yield by about seven percent (+1.24 cubic meters per second [CMS]), from 17.92 CMS under independent operation to 19.16 CMS under reservoir system operation, and delays reservoir depletion, thereby enhancing water supply stability during droughts. The novelty of this study lies in developing a firm-yield estimation method that incorporates reservoir system operation and establishing the concept of a minimum inflow time series. The proposed methodology establishes a quantitative framework for minimum inflow time-series estimation and highlights how reservoir system operation can enhance the resilience and reliability of long-term water supply.
Accurate and timely flood mapping is essential for disaster response, humanitarian coordination, and infrastructure protection. Optical satellite data are often limited during flood events because of cloud cover and heavy rainfall, making Synthetic Aperture Radar (SAR) valuable for all-weather, day-and-night inundation monitoring. This study introduces a deep learning framework for automated flood mapping using Sentinel-1 SAR imagery, based on UNet ++ with a pre-trained EfficientNet-B4 encoder. Flood mapping is formulated as a three-class pixel-wise segmentation problem, distinguishing background, permanent water, and floodwater. The model uses three input channels: VV backscatter, VH backscatter, and a co-registered JRC Global Surface Water permanent water mask. This hydrological prior helps the network separate permanent water from event-driven inundation without requiring additional SAR acquisitions, optical imagery, or terrain data. The model was trained and tested using the Sen1Floods11 benchmark, containing 445 hand-labelled Sentinel-1 tiles from multiple global flood events. It was also evaluated on a held-out Bolivia flood event to assess geographic generalization and as an additional real-world demonstration, the trained model was applied to a Sentinel-1 image acquired over the Koshi Madesh Region flood affected area in Nepal in September 2024. To address class imbalance, a combined Focal and Dice loss was used with AdamW optimisation. The proposed model achieved macro IoU = 0.7340, macro F1 = 0.8294 on Sen1Floods11, and generalised well to Bolivia with macro IoU = 0.6772 and macro F1 = 0.8026. Macro IoU and macro F1 are computed over the permanent water and flood classes, excluding the background class. In Nepal, the model captured the overall water distribution but showed limited strict agreement with UNOSAT, with much of the discrepancy associated with the distinction between permanent and flood water, particularly in the complex braided channels of the Koshi River.
Drought evolution exhibits pronounced multi-temporal-scale variability and coupled linear-nonlinear dynamics. Existing statistical models or deep learning models alone are often unable to achieve satisfactory prediction accuracy across different time scales, while spatial correlation information in regions with complex terrain remains insufficiently exploited. To address these challenges, this study proposes a hybrid forecasting framework integrating the Seasonal Autoregressive Integrated Moving Average (SARIMA) model and the Long Short-Term Memory (LSTM) network to improve the accuracy of multi-scale Standardized Precipitation Evapotranspiration Index (SPEI) forecasting. The proposed model adaptively integrates linear and nonlinear prediction components through a dynamic residual-weighting mechanism and incorporates a lightweight spatial attention module to enhance the representation of spatially correlated features. Using grid-based multi-scale SPEI data collected over the Qinghai-Tibet Plateau during 1981-2022, the proposed framework was trained, validated and tested. Model performance was evaluated across multiple time scales (SPEI-1, SPEI-3, SPEI-6, and SPEI-12) and compared with several benchmark models. Results indicate that the proposed approach consistently outperforms the benchmark models across all considered time scales, with the introduction of the spatial attention module and dynamic residual-weighting mechanism improving the Nash-Sutcliffe Efficiency (NSE) from 0.88 to 0.91. Overall, the proposed framework provides an effective approach for multi-scale drought forecasting in regions with complex terrain.
Wenchuan earthquake-induced landslides generated large volumes of sediment that formed lateral sediment deposits in mountainous river channels, significantly altering flow structures and channel evolution. Under such strongly non-uniform flow conditions, accurate estimation of bed shear stress remains challenging. Because bed shear stress is difficult to measure directly in natural rivers, friction velocity is commonly used as an indirect proxy. This study conducted fixed-bed flume experiments with representative riverbank sediment deposits to evaluate six classical friction-velocity estimation methods under uniform-flow conditions and four methods-the Reynolds shear stress (RSS), turbulent kinetic energy (TKE), vertical turbulent kinetic energy (TKE w '), and logarithmic law methods-under sediment-deposit conditions. The results show that estimated friction velocity is more closely associated with turbulence intensity than with longitudinal mean flow velocity. Turbulence-based methods (RSS, TKE, and TKE w ') showed more consistent responses than the logarithmic law method under strongly non-uniform flow conditions, with the TKE method exhibiting the most stable performance across different flow zones. This study clarifies the applicability and limitations of classical friction velocity estimation methods under different degrees of flow non-uniformity, providing guidance for estimating bed shear stress and conducting sediment-transport studies.
