Study Region: South Sulawesi, Indonesia Study Focus: Machine learning (ML) models have been utilized to reconstruct streamflow data over data-sparse regions, but their aleatoric and epistemic uncertainty have been limitedly assessed. To address this gap, this study focuses on a machine learning-based framework that assesses propagated multiple uncertainties in reconstructing long-term streamflow at 22 stations in South Sulawesi, Indonesia. Three decision-tree and three neural network architectures are integrated via a weighted-stacking ensemble approach, incorporating simulated streamflow by a hydrologic model and ten regional climate indices as predictors. The ensemble performance is evaluated with observational data-related (aleatoric) and model-related (epistemic) uncertainty sources. New Hydrological Insight for The Region: Results show that the decision tree-based ensemble approach outperforms other models, reaching NSE of 0.54 and KGE of 0.48 on median. Incorporating spatial information and large-scale climate indices, such as Pacific Warm Pool Region and Quasi-Biennial Oscillation, improve the predictive performance up to NSE of 0.72 and KGE of 0.68 on median. The aleatoric uncertainty remains as a primary source, driven by inherent variability of streamflow. Results show significant trends only in high streamflow over South Sulawesi from multi-type MK tests, suggesting a high risk of unprecedented floods in a changing climate. This study highlights the potential of AI in reconstructing streamflow data with a comprehensive assessment of multiple uncertainty sources and hydrological variability and trend in data-sparse regions.

North America (NA) has experienced increasingly unprecedented heatwaves, yet distinctions between dry and humid types and their underlying mechanisms remain unclear. Most studies have primarily defined heatwaves based solely on air temperature, while comparatively little attention has been given to approaches that consider soil moisture conditions and atmospheric humidity as intrinsic characteristics of heatwaves. Here, we classify NA heatwaves spatially by exploiting variations in soil moisture and atmospheric humidity, and reveal their dominant drivers. Dry heatwaves in southern NA are characterized by pronounced land-driven processes, intensified by antecedent soil moisture deficits and persistent high-pressure systems. Warming in the North Pacific and North Atlantic further reinforces these conditions. Conversely, humid heatwaves in northwestern NA are primarily driven by atmospheric-driven processes associated with the inflow of warm, moist air and enhanced moisture transport. CMIP6 projections indicate that dry heatwaves will remain dominant but contract spatially in the southern domain under SSP5-8.5, whereas humid heatwaves will intensify continent-wide over NA, and surging nearly 20-fold between the historical (1979-2014) and late 21st-century (2065-2100) period. This accelerated increase leads humid heatwaves to become the prevailing form of late-21st-century summer heatwaves.

Urbanization causes compounding adverse impacts with climate change on regional ecology and society through biophysical processes. Borneo Island in Indonesia has been undergoing rapid urbanization, but the compounding impacts of urbanization under climate change remain unknown. Here, we investigate how idealized urbanization affects regional climate and vegetation using Community Earth System Model version 2 (CESM2) simulations with two land cover scenarios for Borneo: an urbanization scenario and a control scenario with no urbanization for the 2025-2034 period. Results show that urbanization in Borneo alters surface biophysical properties, imbalances the regional surface water budget, and thus leads to a warmer (+0.25 °C) and drier (-0.17 mm/day) climate across the island, with particularly pronounced effects on the dry period climate. Furthermore, these urbanization-induced climate responses contribute to additional vegetation loss. Moreover, vegetation in Kalimantan, the site of Indonesia's new capital, is also sensitive to the combined impacts of urbanization and climate change. This study highlights the importance of considering biophysical climate effects when assessing the compounding impacts of urbanization. This approach can help guide policymakers in updating current climate adaptation plans for sustainable urban development.

Global hydrological models (GHMs) have been used to analyze the hydrological cycle and assess the risk of water scarcity and the hazard of flood and drought hazard worldwide. While it often relies on default parameter values, comprehensive investigations into how parameter definitions in the GHMs impact hydrologic simulations through multi-model and multi-variable analyses remain lacking. Here, we examine how parameter choices affect simulation outcomes in four GHMs—CWatM, PCR-GLOBWB, H08, and HydroPy—by optimizing them using different hydrological variables, including discharge, actual evapotranspiration, soil wetness, and total water storage, across 228 watersheds worldwide. Our results show that GHMs with optimized parameters generally outperform those using default values during the historical baseline period. Default parameters occasionally lead to poor predictions, emphasizing the critical role of parameterization. In particular, default settings produce suboptimal discharge simulations in cold regions, prompting further analysis of parameterization effects on high and low flows. Interestingly, default parameters sometimes overestimate high flows, which could lead to inaccurate flood projections under climate change. On a positive note, the multi-model ensemble approach can enhance the accuracy of projected simulations. However, even within a multi-model ensemble, the use of default parameters may still introduce the risk of overestimating high-flow predictions. This study underscores the importance of continuous improvements in GHMs and highlights the need for a robust parameterization strategy to ensure more reliable climate change projections.

Forecast errors of severe weather events aggravate economic damage and degrade public mental health. Whether forecast errors are underestimated or overestimated can shape public emotional responses differently, which remains unknown. In this study, we investigate the socio-psychological impacts of forecast errors during the landfall of Typhoon Khanun over the Korean Peninsula. We evaluate the predictive performance of multiple lead-hour precipitation forecasts against observational data and conduct a sentimental analysis of over 43,000 online discourses from the NAVER Weather Report Talk platform. Multiple lead-hour precipitation forecasts demonstrate underestimation in the eastern and southeastern regions of the Korean Peninsula and overestimation in the western and southwestern regions. The spatial discrepancies of precipitation forecasts are associated with distinct emotional responses: overestimation (underestimation) makes anxiety and worry (stress and confusion) the dominant emotion types in the discourses from the NAVER Report Talk platform. The findings of this study suggest that expectation-reality mismatch is a key mechanism in risk communication, and the direction of this mismatch differentiates public response. This study provides insights into the potential value of improved forecast accuracy on reducing emotional distress and strengthening public resilience during extreme weather events.

Disaster type dictates how risk communication unfolds on social media, yet the underlying mechanisms remain obscured. Here we conduct a network analysis of over 20,000 Twitter/X users during the 2012 United States Wildfire and Drought and construct an epidemiological model to reproduce information-diffusion dynamics within a social network. We further design targeted intervention simulations perturbating the individual and systematic characteristics. Targeted intervention scenario simulations show that information diffusion is improved primarily by individual characteristics in both drought and wildfire social networks, for example, substantial gain from increased transmission/passing probability and additional gain from reduced recovery probability/forgetting probability. We further find gain in both the extension and speed of diffusion within the centralized wildfire social network, but only the speed within the fragmented drought network. This study identifies disaster-specific pathways for improving social media-based risk communication.

Wildfire variability in Southeastern Australia (SEA) has intensified in recent decades, posing increasing risks to ecosystems and agriculture under a changing climate. However, the mechanisms driving the recent amplification of extreme fire weather remain unclear. Using austral-summer data from 1981-2022, we quantify interannual links between the Forest Fire Danger Index (FFDI) and land-atmosphere variables. Fire Weather Days (FWD) are defined as days exceeding an extreme FFDI threshold each fire season and are validated against satellite-based burned area and fire intensity across SEA. We show that recent fire risk in SEA is characterized not by a gradual increase but by a regime shift in extreme fire weather conditions. An early-2000s transition is marked by enhanced interannual variability and an approximately fivefold increase in FWD, linked to increased positive skewness in daily FFDI. Among FFDI components, the drought factor (DF), representing hydrological stress, exhibits the largest increase in extreme occurrences, especially when co-occurring with high temperature (T) and low relative humidity (RH). The contribution of compound DF & RH & T events to total FWD more than doubles between 1981-2001 (P1) and 2002-2022 (P2). Segmented regression further reveals strengthened interannual FWD sensitivity to DF in P2. In P1, variability reflected atmospheric warming and drying, whereas P2 is characterized by intensified land--atmosphere coupling that amplifies hydrological stress and compound extremes. This transition coincides with changes in large-scale circulation, with the Southern Annular Mode (SAM) emerging as the dominant driver of FWD variability in the recent period, while ENSO exerted a stronger influence earlier. Increased FWD variability is also closely linked to interannual maize yield fluctuations across SEA. These findings highlight a hydrologically-driven regime shift in extreme fire weather and underscore the need for integrated climate-fire-agriculture risk assessment.

Water-related disasters degrade water quality and threaten public health. While their adverse impacts have been well documented, less is known about how associated disturbance alters the public perceived risk. Here, we employ natural language processing to textual data from over 9000 water-related complaint reports of residents in Arkansas over 2008–2024. With water quality data at over 700 monitoring stations, this study examines changes in water quality, complaint behavior, and emotional responses during the 2011 Flood and 2012/13 Drought. Results reveal lower absolute complaint volumes but higher per capita complaint rates in low gross domestic product (GDP) counties despite maintaining relatively good water quality than high GDP counties. During the 2011 Flood and 2012/13 Drought, complaints with the ‘anger’ and ‘disgust’ emotion types increase over low GDP counties while complaints are more objective across counties. We find that flood-driven turbid water, but permissible quality, is associated with changes in complaint behavior patterns of the low GDP, such as topic shifts in the complaints. The results suggest that public complaint behavior is shaped by not only actual water quality conditions, but also other factors such as water-related disaster and socioeconomic structure. This study underscores the need for a state-level public outreach program to help citizens translate complex hydrologic data into actionable knowledge.

Reduced public engagement in water conservation during droughts can exacerbate water scarcity and its associated impacts. However, capturing these coupled human-natural dynamics in hydrologic models remains a persistent challenge. Here, we develop an AI model framework to estimate household water-saving patterns during the 2022-2023 drought in South Korea, integrating both drought conditions and social factors such as news coverage. Results indicate that a spike in drought-related news coverage is associated with higher water-saving rates in metropolitan areas, but not in rural areas. The trained AI model shows that a scenario with an additional 100 news articles per month during the drought onset and caution stages corresponds to an average increase of 13.6 percentage points in model-simulated household water-saving rates in metropolitan areas. This corresponds to potential avoided production costs of KRW 1.14 billion (USD 0.82 million), with more modest gains in rural areas (about 4.0 percentage points and KRW 0.46 billion (USD 0.33 million)). However, the modeled water-saving response in metropolitan areas sharply declines during the drought alert stage when the water saving is already high (above +10%). Our results suggest that drought-related news patterns can serve as a proxy to inform potential region and stage-specific water-saving targets, reflecting differences between metropolitan and rural areas. This study underscores the value of explainable AI in providing a quantitative framework to capture human-nature interactions and support the design of targeted drought communication strategies in a timely manner.

Global warming has profound effects on the terrestrial hydrological cycle, leading to alterations in regional extreme weather patterns. While terrestrial precipitation responses under continued greenhouse gas emissions are well established, the responses of terrestrial precipitation and vegetation feedbacks under climate mitigation scenarios remain uncertain. Here, we investigate terrestrial precipitation changes under idealized negative and zero CO2 emissions scenarios using the Community Earth System Model version 2 (CESM2). Terrestrial precipitation increases by approximately 1.1% at the peak CO2 concentration ( ~ 725 ppm), but shows even greater increases of about 1.9% and 2.5% under substantially lower atmospheric CO2 concentrations following zero ( ~ 600 ppm) and negative ( ~ 430 ppm) CO2 emissions, respectively. Our results suggest that enhanced transpiration from terrestrial vegetation largely contributes to this increase under the negative emissions scenario. Furthermore, despite a substantial increase in terrestrial precipitation, extreme precipitation events and droughts become less severe globally under the negative emissions scenario, even compared to the zero emissions scenario. While near-term mitigation is essential to curb immediate warming, these findings suggest that sustained negative emissions could be effective for achieving long-term reductions in hydrological extremes and enhancing terrestrial water availability.

Accurate reservoir outflow simulation is crucial for modeling streamflow in reservoir-regulated basins. In this study, we introduce a knowledge-guided Long Short-Term Memory model (KG-LSTM) to simulate the outflow of reservoirs-Fengshuba, Xinfengjiang, and Baipenzhu in the Dongjiang River Basin, China. KG-LSTM is built on the standard hyperparameters-optimized-LSTM and the loss function considering reservoir operation knowledge. Model uncertainty is analyzed using the bootstrap method. We then propose a hybrid approach that combines KG-LSTM with the Three-parameter monthly hydrological Model based on the Proportionality Hypothesis (KGLSTM-TMPH) for streamflow simulation. The propagation of inflow errors to outflow simulations is studied across the three reservoirs. Results show that KG-LSTM enhances accuracy and reduces uncertainty in outflow simulations for three reservoirs compared to LSTM, particularly for the multi-year regulated Xinfengjiang Reservoir: KG-LSTM improves Nash-Sutcliffe efficiency (NSE) from 0.59 to 0.64, reduces root mean squared error (RMSE) from 55.59 m3/s to 54.84 m3/s, and decreases the uncertainty index relative width (RW) from 0.55 to 0.51 during the testing period. For streamflow simulations at four downstream hydrological stations, the hybrid model KG-LSTM-TMPH achieves NSE values above 0.87 and outperforms LSTM-TMPH, particularly in the dry season. Inflow errors impact outflow most significantly for the Xinfengjiang Reservoir in April and May, for the Fengshuba Reservoir throughout the year, and for the Baipenzhu Reservoir in July and August. This study enhances reservoir outflow modeling by integrating reservoir operation knowledge with deep learning. The hybrid KG-LSTM-TMPH approach shows practical potential for streamflow simulation in reservoir-regulated basins, offering valuable applications for water resource management.

A 9-km daily gridded streamflow dataset is generated using the Variable Infiltration Capacity-River Routing Model (VIC-RRM) across the Ganges-Brahmaputra-Meghna River basins over 1951-2023, forced by the ERA5-Land reanalysis data for naturalized streamflow. Physically consistent streamflow forecast data is also generated forced by the ECMWF S2S forecasts. The performance of the dataset is evaluated using observed streamflow data from three gauge stations in Bangladesh along the streams of Ganges, Brahmaputra, and Meghna Rivers, calculating the modified Kling-Gupta Efficiency (mKGE) metric for the 365-day climatology. For Ganges, Brahmaputra, and Meghna Rivers, the mKGE values of reconstructed streamflow data are 0.50, 0.75, and 0.25, respectively. Comparing with the reconstructed streamflow data, the streamflow forecasts show a good agreement with mKGE values of 1.00, 0.97, and 0.91 at three gauge stations, respectively. This dataset provides physically consistent reconstructed and forecasted streamflow data at high resolution, offering a valuable resource for the assessment of climate variability and change and the development of river basin-specific water management strategies in the Ganges-Brahmaputra-Meghna Rivers in Bangladesh.

Social network plays a critical role in risk communication diffusing information in near real time. Disaster-affected communities utilize their social network to report catastrophic damages and increase the perceived risk of the ongoing disaster by non-affected communities, which enhance their willingness to donate and support emergency aids to the affected communities. Previous studies have focused on social network structure or information diffusion separately. This study strives to reproduce the social response to natural disasters aims integrating the two aspects of social network structure and information diffusion. This study focuses on two classical and catastrophic U.S. disasters, such as 2012 flash drought and wildfire, to establish the social network during these two disasters and understand difference in the patterns of the risk communication within the data-driven social network and random social network (e.g., (the equal chance/importance of a nodes). Random social network is made from the LFR benchmark algorithm using the properties of the data-driven network, including node number, degree distribution, community distribution, and average degree. This study leverages over 120,000 (53,000) tweets that contains a term, drought (wildfire). In this study, a Susceptible-Infected-Recovered (SIR) model is employed to simulate the information diffusion patterns using the data-driven and random social network. After fitting SIR model with the Twitter data using these two social network-based simulations, this study aims to assess 1) the impact of the structure difference on risk communication and 2) the impact of influential users in different social network structures. Result shows that the trained SIR model using the data-driven social network reproduced the observed information diffusion patterns for the 2012 drought and wildfires but with relatively higher uncertainty in the information diffusion pattern for wildfires. The SIR model simulation with data-driven social network shows a faster information diffusion pattern with a higher information reach rate than that with the random social network. In closing, this study discusses limitations and opportunities of next-generation social dynamic modeling for natural disaster risk communication. This study highlights the value of an interdisciplinary approach in improving risk communication and developing a more efficient and effective mitigation policies for not only droughts and wildfires and other natural disasters.

Long-term streamflow data at a hyper-resolution (less than 1 km) is essential for hydroclimatic extreme and ecological assessment, which is not available over a river basin where rapid socioeconomic growth have been experienced. Here, we use the Variable Infiltration Capacity-River Routing Model (VIC-RRM) to reconstruct naturalized daily streamflow at 90–meter resolution for the Geum River, one of South Korea’s major rivers, over 1951–2020. VIC-RRM demonstrates high temporal consistency with a correlation coefficient exceeding 0.6 for observed streamflow seasonality at over 60

In this study, we develop a low-dimensional recursive model using deep learning (DL) to understand the dynamics of the El Ni & ntilde;o-Southern Oscillation (ENSO). Unlike most existing research that relies on Coupled General Circulation Models (CGCMs), we explore a DL technique as an alternative approach to simulate ENSO characteristics. To replicate the observed stochastically excited oscillations, we incorporate stochastic noise into the recursive process of the DL model. Our long-term simulations demonstrate that the DL model effectively reproduces ENSO characteristics comparable to those captured by CGCMs. Additionally, we conduct experiments to analyze the interactions between ENSO and the Indian and Atlantic Oceans, evaluating their impacts on ENSO dynamics. Beyond capturing ENSO characteristics, the DL model exhibits skillful ENSO prediction capabilities. Using eXplainable AI (XAI) methods, we identify the contributions of each variable to ENSO predictability. Our findings suggest that this DL model serves as a valuable tool for understanding climate dynamics at a relatively low computational cost, providing an alternative to complex physically-based models.

Land use change (LUC) and climate are major factors constraining terrestrial ecosystem. As these impacts are expected to intensify in the future, it is necessary to quantify the effects of anthropogenic LUC on the terrestrial biosphere and its interactions with regional climate. Here, we identify abrupt decreases in the Leaf Area Index (LAI) from 2015 to 2100, primarily driven by LUC through cropland expansion. We further examine the role of climate change in the identified LAI reductions. In areas experiencing LUC, a negative relationship between LAI and surface temperature leads to a positive feedback, which reinforces decreasing LAI and increasing warming. Moreover, prolonged regional dry conditions and a shift towards drier climate conditions further exacerbate LAI reductions. These climate change impacts have a large spatial variation and cause western South America to show the most pronounced LAI sensitivity to LUC under the shared socioeconomic pathways (SSP) 3-7.0 scenario. These results highlight the combined effects of LUC and climate change, leading to future abrupt decreases in vegetation cover.

Land‐falling tropical cyclones (TCs) bring much‐needed precipitation for drought mitigation, which veils the attribution of missed TC landfalls to drought. Here we assess the global impact of TC landfalls on soil moisture droughts over 1980–2020 from offline hydrologic model experimental simulations forced by precipitation with and without TCs. Results demonstrate that the absence of TC precipitation causes severe soil moisture droughts with different precipitation‐soil moisture‐streamflow pathways. Over arid/semi‐arid regions like Oceania, soil moisture from TC precipitation is desiccated within a year, that is, the absence of TCs triggers severe droughts while soil moisture is not completely dried out even without TC precipitation over humid regions like East Asia. This study highlights the diverse impacts of missed TC landfalls on soil moisture droughts.

A severe drought causes catastrophic economic losses, resulting in mental degradation/deterioration. While drought monitoring has been focused on detecting and characterizing an emerging drought (physical system-focused), artificial intelligence with big data from social monitoring provides a unique opportunity to investigate sentimental alterations of the public along the drought propagations and explore their triggers (social system-focused). This study examines the potential of an AI technique, Natural Language Processing (NLP), in monitoring sentimental alterations of the public. This study is a case study of the recent Korea drought, leveraging X (formerly, twitter) and Google Trends data. In this study, we evaluate the seasonal-to-seasonal predictability of drought measures in the southwest region of the Korean Peninsula for the 2022/23 period and analyze spatiotemporal changes in media and public interest in drought phenomena during the 2022/23 drought period through newspapers and social media. Initially, to understand the predictability of drought measures in March 2023, we evaluate the predictability of drought measures based on probabilistic and deterministic seasonal-to-seasonal forecasts for the 2022/23 Korea drought. Subsequently, using drought-related articles in newspapers and Twitter data from the 2022 to 2023, we utilize natural language processing and text mining technologies to detect and monitor the topic and emotional alternation of the titles of news article and public, respectively, regarding the 2022/23 drought. The results of this study indicate that the predictability of drought measures in May 2023 is skillful at the sub-seasonal scale but limited at the seasonal scale. Statistical forecasts provide crucial information on precipitation needed for drought recovery through weekly precipitation forecasts, aiding in assessing the drought condition. The public interest in the 2022/23 drought shows spatiotemporal differences based on the drought-affected areas and drought stages. Especially in April 2023, when a severe drought occurred in the southwestern Korean region, an increase in the number of newspaper articles with negative titles was observed, and negative emotions were detected from the social media data. This study provides an insight about the role of AI in developing the next-generation drought monitoring system.

Analytic Hierarchy Process (AHP) of pluvial flood risk assessment has been widely used, incorporating multiple assessment indices. However, uncertainty assessment of expert judgement-based flood risk remains limited. This study proposes a Machine Learning (ML) model-based AHP approach, using pluvial flood-related data of hazard, exposure, vulnerability, and capacity of South Korea over 2002-2021. In this study, the trained eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) models successfully predict flood economic losses using 21 flood-related variables, outperforming LightGBM and CatBoost. Permutation importance scores from the trained XGBoost and RF models are used to estimate the mean and 95 % confidence intervals of the assessment factor weights. Both models show that rainfall amount, river area, population density, and green belt area are important factors for flood damage prediction, but the XGBoost (RF) model identifies impermeable areal fraction (river area) as the most important component in exposure, resulting in disparity in the uncertainty range in major cities over South Korea where the XGBoost and RF models show a high risk consistently. This study substantiates the practical application of the proposed ML based-AHP approach for uncertainty assessment of flood risk, highlighting the need for balanced land development and green infrastructure for flood mitigation.

Hydrologic projections of current regional climate models are uncertain over mountainous regions because of the poor representation of complex terrains. Here, we use multiple AI models to generate observation-constrained projections of the occurrence probability of record events (REs) of annual maximum and minimum daily streamflow ( Q _max and Q _min ) in Pakistan. First, we trained and tested the AI models, using the occurrence probability of REs from simulated (input) data from five regional climate models and observed (output) river discharge data from 1962–2022. Then, we use the trained AI models to project the observation-constrained occurrence probability of REs until 2099 and conduct principal component analysis for the robust selection of AI model ensembles. The observation-constrained projections detect more REs in Q _max and Q _min in the late 21st century than expected under the stationary system does, highlighting intensifying hydroclimatic extremes. The upper Indus River shows that the return period of REs in Q _max and Q _min is approximately 15 years. The Chenab and Kabul Rivers (Jehlum River) are prone to more REs in Q _max ( Q _min ), with a return period of approximately 11 years. Uncertainties in the RE occurrence probability projections from regional climate models can be attributed to the heterogeneity in hydroclimatic and physiographic characteristics across the study basins. This study suggests the nonstationarity of four major river basins in Pakistan, where water resources management should be updated with basin-specific strategic plans.