Research

Traditional dynamic or statistical downscaling techniques have only modestly improved Indian summer monsoon rainfall simulations. While deep learning models like YNet have shown promise in global downscaling efforts, we found their application to Indian Summer monsoon rainfall (ISMR) has been less effective, particularly for extreme rainfall and intraseasonal variability. We developed two advanced super-resolution deep learning models to overcome these limitations: YNet_D and YNet_DE. YNet_D enhances the downscaling accuracy by integrating atmospheric variables, informed by monsoon dynamics, with coarse precipitation data as predictors. YNet_DE prioritizes extreme rainfall using a weighted loss function, effectively addressing the challenge of simulating extremes. 

A. In terms of extremes, YNet_D and YNet_DE reduced the bias in the precipitation greater than 95th percentile (R95p) from 78.27 mm with YNet to 56.52 mm and 40.22 mm, respectively.

B., C. Additionally, both models showed significant improvements in simulating ISMR’s intraseasonal and interannual variations 

Our models demonstrate strong transferability to General Circulation Models (GCMs), with YNet_DE notably lowering the multi-model mean bias for R95p from 114.3 mm in the original GCMs and 130.0 mm in YNet simulations to 64.4 mm. Unlike conventional statistical methods, these models retain the critical physical dynamics of ISMR, offering a robust solution that preserves the system’s variability and complexity. For details check this link.

Quantile mapping based bias correction and spatial disaggregation (BCSD) have emerged as the de facto standard for rectifying bias and scale-mismatch in global climate models (GCMs) leading to novel climate science insights and new information for impacts and adaptation. Focusing on critical variables crucial for understanding climate dynamics in India and the United States, our evaluation challenges the premise of BCSD approach. 

A. We find that BCSD corrects randomly generated fields to match their statistical properties to observations which shows that BCSD overcorrects GCM simulations to observed patterns while minimizing or even nullifying science-informed projections generated by GCMs.

B. We also show that BCSD incorrectly captures complex complex climate signals even with the trend-preserving bias correction.

C. Our evaluation in the context of the Walker circulation shows inability of QM and its multivariate derivatives to adequately capture multivariate and spatial-temporal dependence patterns.

For details on this work visit here.

The 10% coefficient of variation (COV) in All India Monsoon Rainfall (AIMR) assumes paramount importance as the government classifies years with AIMR above 110% as surplus and below 90% as deficit, shaping critical decisions. While most Indian regions exhibit a 20-40% COV for grid-level monsoon rainfall, the mystery surrounding AIMR's 10% COV persists. 

A. An examination of 40 years of gridded monsoon rainfall data exposes the negative covariance between Central India (CI) and Northeast India (NE) constraining the COV of AIMR to approximately 10%.

B. Utilizing Lagrangian backtracking of moisture in reanalysis data, we found that during CI's deficit rainfall years, NE experiences a surplus with 61.69% moisture from terrestrial sources. In CI's surplus years, a major fraction of the moisture supply from the ocean is confined to CI, resulting in NE facing deficit rainfall. 

Our findings emphasize the non-intuitive processes leading to the 10% COV in AIMR cautioning against relying on this value for scientific studies or planning. For details visit here.

Assessment of the response of the Himalayan river flows to climate change is complex due to multiple contributors: rainfall, snowmelt, and glacier-melt. 

A. We integrated a glacier-melt model with VIC and validate the model output with observed streamflow in five river basins in the Himalayas, at daily scale. 

B. We disentangled the different components of streamflow in Himalayan rivers and quantified the contribution of each to the total stream flow.

C. Climate model simulations show a decrease in the spring onset times and decreases in the center of volume of streamflow, which suggests that there will be increased flows in the early part of year and reduced flows later in the year for both RCP4.5 and RCP8.5 scenarios. 

For details and results for other Himalayan basins in the study visit this link.