| Sl.No | Name | Title | Abstract |
| 1 | Aswin K | A Comparative Study Of Hec-Hms And Machine Learning Models For Streamflow Projection And Evaluating Climate Change Impacts On The Chaliyar River Basin, India | The Accurate simulation of streamflow is crucial for effective water resource management, in region like Chaliyar River basin (CRB), and the basin witnessed drastic changes in streamflow in recent years. This study evaluates the efficiency of six hydrological models, HEC HMS, ANN, SVR, RFR, GBR, and MLR in predicting streamflow over CRB. A comparative analysis has been performed between these models with the help of various performance metrics such as the R2, NSE, PBIAS, and RMSE. The best performing model was used to project streamflow up to 2060, using CMIP6 projected data of precipitation and temperature. The results showing that, all ML models outperformed HEC-HMS, with ANN showing the best performance (R² = 0.92, NSE = 0.90), and was used to project future streamflow. As a conventional process oriented hydrological model, HEC-HMS achieved R² = 0.72 and NSE = 0.64. ANN successfully captured major peaks, including the 2019 flood (4164.7 cumecs), but all models failed to simulate the 2024 flood peak. Bias-corrected meteorological inputs from three GCMs were used, with ESN1-C ensemble mean performing best (R² = 0.81 for precipitation, 0.96 for temperature). Future projections under SSP2-4.5 indicated a peak discharge of 3039 cumecs in 2029 and highlighted extreme reduction in streamflow in years like 2026, 2040, 2042, and 2043, along with two moderate flooding conditions in 2035. Streamflow in the CRB is projected to decline, especially after 2055, due to rising temperatures, altered rainfall pattern. Despite decent rainfall in years like 2034 and 2059, streamflow remains relatively low, suggesting the influence of factors other than rainfall. The ANN model performed well in simulating future changes, showing the value of ML and bias-corrected data for managing water resources under climate change. The novelty of this research lies in the advanced modelling approach for accurate projection of streamflow for a small river basin, suffering from data scarcity. |
| 2 | Laxmi Pandey | Machine Learning based modeling of groundwater salinity- a case study on Israel | Increasing salinity and pollution has been a serious issue causing degradation in water quality and agricultural productivity. Thus, for a comprehensive understanding of its underlying causable factors and to identify important drivers of salinization, we have used wide range of datasets in this study. We have effectively integrated different datasets, including climatic, topographic, soil, and anthropogenic covariates, to create a robust framework for machine learning based predictive models of groundwater salinity across Israel's seven hydrological basins. Key findings from the study include the superior performance of Random Forest (RF) and Gradient Boosting (GB) models over logistic regression for classification tasks, while the Long Short-Term Memory (LSTM) model is found to provide better results compared to that of Feed Forward Neural Network (FFNN) in regression tasks. Additionally, feature importance analysis using Recursive Feature Elimination (RFE) with Random Forest methods yielded significant insights into the covariates such as Shoreline distance, Elevation, Area, total wetness index (TWI), Aridity Index, Population density, Water extraction, temperature, Precipitation and soil type, influencing groundwater salinity. We also carried out causality analysis via Double machine learning using various predictive models. This approach allowed us to shortlist the few predictors which may have strong causal impact on groundwater salinity after removing the confounding effects. The predictors showing strong causal relation with the predictand groundwater salinity included Temperature, TWI, land use land cover, Elevation, Count agricultural factories, Precipitation, Area, Water extraction, Population density, Aridity index. Overall, this approach highlights the variability in groundwater characteristics across different regions, emphasizing the need for tailored strategies to address salinity issues effectively. |
| 3 | Manali Saha | Precursors of moist heatwaves over Indian Subcontinent | Heatwaves represent one of the most hazardous meteorological phenomena, posing significant risks to millions of individuals. Defined by prolonged periods of extreme temperatures, these events exert considerable effects on ecosystems, economies, and human mortality rates. When combined with high humidity, these events induce substantial heat stress across the affected regions. India, as a prominent hotspot, experiences heatwaves during the pre-monsoon season, which are associated with both moist and dry mechanisms. Moist heatwaves, characterized by elevated wet bulb temperatures, result in high fatality rates among humans and mammals. Despite the high population density and the context of climate change, the origins of these moist heatwaves have not been thoroughly investigated until now. In this study, we examine the precursors of moist heatwaves in the Indo-Gangetic Plains using the Eulerian temperature decomposition equation to identify the dominant processes responsible for the formation of these events. Previous literature suggests that advection is the primary component in triggering these events; however, our analysis demonstrates that the effect of advection is minimal due to the weak temperature gradient in the tropics. To investigate the precursors, we extend our analysis from the pre-heatwave period to the onset of the heatwaves. Our findings indicate that pre-monsoon showers are responsible for the formation of moist heatwaves. These showers are associated with nighttime low-level clouds that trap outgoing long-wave radiation, thereby accumulating heat content and causing temperatures to rise. Furthermore, these rainfall activities must be supported by mid-tropospheric dryness (MTD) to sustain throughout the period. The MTD aids in maintaining the low-level clouds resulting from shallow convection and does not promote deep convection. We emphasize the importance of local atmospheric conditions, along with large-scale activities (that trigger anticyclones in the upper troposphere), in sustaining the intensity of heatwaves. The findings of this study will contribute to the development of heatwave early warning systems at localized scales. |
| 4 | Manu Jangid | Forecasting Of Lightning Density Over Northern Part Of India Using Machine Learning |
Machine learning is a strong method for predicting lightning activity by learning patterns from historical data. This study focuses on the prediction of lightning density using machine learning techniques over Northern India. Models were trained using LIS/OTD 2.5° Low-Resolution Monthly Climatology Time Series (LRMTS) lightning data as the target variable, with K-index, Convective Available Potential Energy (CAPE), Surface Sensible Heat Flux (SSHF), and 2-meter air temperature (T2M) as predictor variables for the period 2000-2011 and tested for 2012-2013. Three machine learning models, namely Decision Tree, Random Forest, and XGBoost, were developed separately for the monthly dataset using k-fold cross-validation with each year treated as one fold to ensure temporal robustness. Additionally, a Long Short-Term Memory (LSTM) model was designed specifically for the daily dataset to capture short-term temporal dependencies. Among all models, XGBoost showed the best performance on monthly data with an R² of 0.84 and a Mean Absolute Error (MAE) of 0.0113. The LSTM model demonstrated reasonable accuracy on daily data with an R² of 0.62 and an MAE of 0.0169. Spatial and temporal evaluations revealed that the models could effectively replicate observed lightning patterns, though slight underestimations occurred in regions with intense lightning activity. Future research may explore hybrid strategies that combine Numerical Weather Prediction (NWP) models with machine learning to enhance lightning forecasting throughout India. |
| 5 | Naveen Kumar E | Machine Learning for Intensity and Track Prediction of Monsoon Low Pressure Systems | Monsoon low-pressure systems (LPS) are pivotal components of the South Asian monsoon, contributing significantly to regional hydrology and agriculture while also causing catastrophic extreme rainfall events. Current numerical weather prediction (NWP) models have limitations in reliably forecasting the evolution of LPS, including systematic track errors, intensity biases, mis-timing of peak intensity, and failures of modeled low pressure areas to intensify into depressions as observed, resulting in only moderate skill for LPS evolution even in short-range forecasts of 1–3 days. This research addresses this critical gap by developing a deep learning approach for LPS forecasting. We implemented a Temporal Fusion Transformer (TFT) architecture to forecast both LPS intensity and position. By leveraging the ability of TFT to capture temporal relationships at multiple scales, the model could achieve accurate multi-horizon predictions of LPS behavior. The TFT framework was trained on carefully selected meteorological variables, with various experimental configurations, testing the impact of input duration (12 vs. 24 hours) and inclusion of background environmental conditions. Our results demonstrate that the TFT achieves near-zero bias in intensity forecasting and position errors competitive with or better than current NWP models, while requiring far fewer computational resources. Notably, different configurations exhibited distinct strengths: the 24-hour with background setup achieved the highest monsoon depression hit ratio (78.7%) and strongest intensity correlation (R² = 0.923), while the 24-hour without background configuration demonstrated superior track prediction at longer lead times. Remarkably, the model demonstrates strong performance even with minimal input data of just 12 hours, suggesting that it effectively captures early atmospheric signals and initial tendencies that are indicative of subsequent LPS intensification. Furthermore, the model accurately predicts landfall timing and position, offering valuable lead time for disaster management. By enabling more rapid and resource-efficient operational implementation, this approach could substantially improve early warning systems and disaster preparedness in regions vulnerable to these high-impact weather systems, potentially saving lives and reducing economic losses. |
| 6 | Rahul Kashyap | Growing ecological implications of moisture stress in Indian during Indian Summer Monsoon reveals Machine Learning | Current definitions of drought, which focus on its impacts on agriculture, water resources and society have a limited perspective as it overlooks its ecological impacts. Ecological drought refers to the hindered growth and development of vegetation, which has negative effects on the ecosystem such as altered carbon, nutrient and water cycling. India exhibits strong carbon-water cycle connection and is an agrarian economy. We investigate ecological droughts, its underlying causes and their impact on vegetation dynamics in India during the Indian summer monsoon (ISM) period for the past two decades (2000–2019). We find that the ecological droughts are rising in most parts (except western) of India due to enhanced land evaporative, meteorological and atmospheric aridity. The Machine Learning (ML) based Random Forest (RF) algorithm suggest meteorological aridity (23.9%) and ocean warming (18.2%) largely drives the ecological droughts in India during ISM. Causal analysis reveals ocean warming indirectly triggers ecological droughts in India, as it affects its other drivers. The rising ecological droughts leads to browning during ISM, which is a concern for sustainability, food security and climate change mitigation. There is a need to consider the ecological implications of droughts in policies and combat its threats such as the prevention of crop failures, famines, degeneration and fragmentation of forests, and socio-economic issues. It may be feasible to alleviate ecological droughts by modified policies and efficient water management. Employing conventional solutions are inadequate as these approaches are mostly lacking in drought planning, and their effectiveness and cost are seldom measured or compared to infrastructure-based mitigation strategies. Hence, it is imperative to incorporate ecosystem services and vulnerability assessment into the planning process to effectively tackle ecological droughts. Proactive resource management strategy such as forest thinning and nature-based solutions that align with natural processes can effectively mitigate ecological drought vulnerability. The findings of the study will help in effective planning for mitigation and adaption of the adverse impacts of droughts on ecosystems in India and is applicable to similar bioclimatic regions of the world. |
| 7 | Shrutee Jalan | Representation of Tropical Intraseasonal Variability in AI Models | Intraseasonal variability is a prominent component of variability in the tropical atmosphere. Key modes of tropical intraseasonal variability (TISV), such as Convectively Coupled Equatorial Waves (CCEW) and the Madden-Julian Oscillation (MJO), present significant forecasting challenges due to their nonlinear and multiscale nature. The increasing demand for accurate and timely weather forecasts has driven the integration of AI into Numerical Weather Prediction, particularly for short to medium-range weather forecasting. Leveraging AI models to improve the representation and prediction of TISV holds promise for enhancing the accuracy of extended-range forecasts, particularly in the tropics, which remains a significant challenge due to the complex variability of the tropical atmosphere. In this study, we evaluate the representation of TISV modes using three state-of-the-art AI models — PanguWeather, GraphCast, and FourCastNet V2. Wheeler-Kiladis plots generated from the model outputs show some power in the Kelvin wave and Rossby wave domains, along with strong power in the MJO band. To assess whether the power captured by the models reflects physically meaningful structures, we analyse the representation of Kelvin waves, Rossby waves, and the MJO in the models. For Kelvin waves, all the three models exhibit eastward-propagating disturbances between 5°N and 5°S, with strong zonal wind patterns resembling Kelvin wave disturbances but phase velocities in the range of 4–7 m/s. The models generally reproduce the tilted vertical structure of Kelvin waves in at least one variable, such as specific humidity or temperature anomaly, with the magnitudes of anomalies across variables consistent with documented studies. For Rossby waves, all the models capture westward-propagating gyres around 10°N/S, with phase speeds of approximately 5–6 m/s, aligning with documented value. Most models reproduce the correct phase of anomalies in at least one variable, although some fail to capture the expected vertically upright structure. For the MJO, all three models exhibit a large-scale eastward-propagating disturbance with strong zonal flow. However, none of the models capture the gyres typically associated with the MJO. Thus, our results suggest that these models do capture some of the physical characteristics of the TISV modes examined in this study. Future work includes creating a composite of events to obtain more robust results and case study analyses to evaluate how these models compare with real-world observations. |
| 8 | Sreyasi Biswas | Precipitation downscaling using Kernel machines | Precipitation downscaling using Kernel machines Sreyasi Biswas1, Parthe Pandit2, Vishal Dixit1 1Centre for Climate Studies, Indian Institute of Technology, Bombay, Mumbai, India 2Centre for Machine Intelligence and Data Science, Indian Institute of Technology, Bombay, Mumbai, India Abstract Satellite-remotely sensed precipitation provides uniform data over a broad spatiotemporal extent; nevertheless, it is an indirect measure, typically impacted by retrieval algorithm uncertainties. On the contrary, the ground-based rain gauge data offers historic, point-wise, and direct measures of precipitation, yet their heterogeneous spatiotemporal distribution, and consequently their spatial interpolation techniques introduce biases. Such complementary strengths of both sources motivate the integration of the high spatiotemporal satellite data and ground-based rain gauge data to improve precipitation estimates overall. We have employed a kernel machine to refine satellite remotely sensed Integrated MultisatellitE Retrievals for GPM (IMERG) (0.10o x 0.10o) precipitation data using the India Meteorological Department (IMD) (0.25o x 0.25o) precipitation data. Owing to IMERG being standardized with monthly rain gauge data, we identified regions with high monthly correlation coefficient (r) values. Thereafter, we trained the kernel machine to enhance IMERG by employing an appropriate scaling factor, with more weightings to the highly correlated regions initially identified. The preliminary results showed that when the kernel machine is trained on r ¬> 0.65 regions, the magnitude of Mean Absolute Error (MAE) and the outliers are relatively lesser than of the native IMERG. The ongoing efforts focus on optimizing the hyperparameters of the kernel machine for more precise estimation. Keywords: Downscaling, precipitation, kernel machine, IMERG, IMD |
| 9 | Swadesh Mohapatra | Assessment of Surface Urban Heat Island over Bengaluru City in India | Assessment of Surface Urban Heat Island over Bengaluru City in India Swadesh Mohapatra* and Krushna Chandra Gouda 1CSIR Fourth Paradigm Institute, Wind Tunnel Road, Bengaluru, India 2Academy of Scientific and innovative Research, Ghaziabad, UP, India *e-mail: swadesh.fpi24j@acsir.res.in The population approaching 14 million in the Bengaluru's metropolitan area in South India and is grappling with various environmental challenges like poor urban planning, including unchecked urbanization, air pollution, water scarcity, and waste management issues etc. The impact of climate change (CC) is also well observed in the urban Bengaluru resulting in the local Urban Heat Island (UHI). The interaction between local UHI and global CC creates challenges to human health, wellbeing and development. This study uses MODIS-Aqua Land Surface Temperature (LST) data for a decade (i.e., 2015-2024) to examine the UHI effect over the city. Climatological analysis of night time LST shows an average annual temperature-increasing trend between the urban Bengaluru and its neighboring suburbs and villages. This difference is computed at monthly scale and the fluctuations are being estimated using the satellite and validated against the ground observations. The Land use Land cover estimation are also linked to the UHI effect and the role of vegetation cover in the LST distribution is also quantified and it indicates the direct impact. This study will help in understanding the LST dynamics in the UHI effect over a rapidly urbanization city and can be used in the climate projection studies offering a ways to guide the urban planners, disaster managers and policy makers. |
| 10 | Triparna Sanyal | Dynamic Mode Decomposition to study ocean vertical mixing | Fluid mixing in world's oceans is a fundamental physical process that governs the redistribution of heat, salinity, carbon, nutrients, and momentum throughout the water column. This phenomenon is highly nonlinear in space and time and exhibits broad range of scales of motion.The existing global circulation models usually parameterize effects from these unresolved scales however there remains persistent uncertainties in such traditional physics-based parameterization schemes. Adopting data-driven methods can help us overcome challenges from improper representation of key physical processes by providing computational efficiency and generalization. In this study, we explore a technique known as the Dynamic Mode Decomposition (DMD) which is a dimensionality reduction technique. Given a set of spatio–temporal data, DMD extracts coherent patterns from the data and reduces the system’s complex evolution into its dominant features and essential components. We import the vertical profiles of temperature, salinity from CMIP6 model data and extract the dominant modes of variability using DMD to predict future ocean stratification. |
| 11 | Usman Hyder Patoo | Downscaling of Satellite Soil Moisture for field scale predictions | Efficient water resource management is critical for addressing the growing challenges of water scarcity and agricultural sustainability, particularly in regions with significant climatic and anthropogenic pressures. This study focuses on downscaling satellite-derived soil moisture (SM) data to farm-scale root zone soil moisture (RZSM) to support precision agriculture and optimise irrigation practices. Current satellite datasets, such as SMAP, provide coarse-resolution soil moisture data that fail to capture farm-scale spatial variability influenced by factors like soil heterogeneity, topography, and human activities. To address this gap, we aim to employ machine learning techniques and physics-based modelling to develop high-resolution soil moisture prediction models. Validation of the satellite data indicated its inability to capture farm-scale variability. Our preliminary efforts utilised an LSTM-based approach to downscale RZSM using site-level sensor data from June 2023 to January 2025. The model achieved good results, including an RMSE of 0.01 and R² of 0.88. To enhance model robustness, we aim to incorporate multiple predictors, such as precipitation, vegetation indices, and land surface temperature, and use robust machine learning architectures to capture spatiotemporal heterogeneity effectively. |
| 12 | Varsha Ganguly | Classification of Clouds over Rajasthan for Insights into Desert Cloud Formation and Evolution | Clouds play a very important part in Earth’s climate system; they regulate the planet’s energy balance, atmospheric dynamics and hydrological cycle. They remain one of the most challenging meteorological phenomena to observe, study or forecast due to their dynamic nature and quickly fluctuating nature in both spatial and temporal scales. Changes in cloud properties can drive changes in precipitation and climate patterns, which is very essential for areas like Rajasthan, which have an arid and semi–arid climate. This study aims to examine the variability of cloud characteristics over Rajasthan, India, during the southwest monsoon months (JJAS) for insights into the cloud properties of the region. Datasets from NOAA and MODIS were used to examine the cloud properties and further calculate the cloud cover probability of a region. The results reveal significant spatial heterogeneity in cloud cover probability over Rajasthan, with western districts showing the lowest cloud presence. The dynamic classification further highlighted the dominance of low and mid-level cloud types, consistent with the arid and semi-arid climatic profile. Results imply a strong rainfall variability and point to a strong coupling between cloud microphysical characteristics and regional precipitation processes. Overall, this study emphasises the critical role of cloud dynamics in shaping monsoon behaviour over desert regions in Rajasthan and underlines the importance of high-resolution satellite datasets and classification techniques in improving our understanding of cloud–climate interactions in vulnerable regions. These insights can inform future efforts in climate modelling, weather prediction, and water resource planning for drought-prone areas. |
| 13 | Yogenraj Pravin Patil | Climate Network Analysis of the Large-scale Onset and Withdrawal of Indian Monsoon | With summer season rainfall accounting for about 80% of annual rainfall, the food security of South Asia hinges on the Indian summer monsoon season rainfall delineated by onset and demise. While the Indian monsoon is a planetary-scale phenomenon, an objective definition of planetary-scale onset, demise, and the rainy season is lacking. The operational definitions currently do not differentiate local-onsets from large-scale ones making their predictions inutile to the farmers. Using complex networks, our new definition identifies large-scale onset at each location when the local-onset becomes part of the large-scale cluster of onsets. We establish that monsoon onset is a phase transition characterized by two abrupt jumps and a consistent northward propagation of the rainfall band follows the second jump. Contrary to the conventional wisdom, we show that the Indian monsoon establishes first over Northeast India followed by that over the Indian peninsula through two abrupt growths in the largest cluster of onsets. |