Machine Learning–Based Attribution of Hydroclimatic Extremes and Agricultural Yield Risk in the Brahmaputra Basin, Assam, India under CMIP6 Scenarios

Changing climatic conditions have led to hydroclimatic extremes, posing significant risks to water availability, agricultural productivity, and food security in climate-sensitive regions. The Brahmaputra River basin, situated in northeastern India, largely within the state of Assam,  is particularly vulnerable to climate change, as rain-fed rice cultivation in this area is highly dependent on the monsoon.This study assesses historical and projected climate-yield relationships at the district level in Assam using a machine learning framework.

The present analysis utilizes hourly ECMWF ERA5 surface-level data (AgERA5), which includes agrometeorological variables such as 2 m temperature, total precipitation, and reference evapotranspiration. Agricultural drought stress has been evaluated using the Standardized Precipitation–Evapotranspiration Index (SPEI), sourced from the global SPEI database. The Expert Team on Climate Change Detection and Indices (ETCCDI) indices were employed to evaluate climate extremes, including various temperature indices (annual maximum and minimum of daily maximum and minimum temperatures: TXX, TXN, TNX, TNN), diurnal temperature range (DTR), and precipitation extremes (maximum 1-day and 5-day precipitation amounts: RX1day, RX5day).

These indices were temporally correlated with district-level rice yield data and spatially aggregated across the Upper, Middle, and Lower Brahmaputra Basin regions. Long Short-Term Memory (LSTM) neural networks were applied to capture the nonlinear and temporal relationships between agrometeorological variables, climate extremes, and rice yield variability. To account for model uncertainty, multi-model ensemble spreads from CMIP6 projections under SSP2-4.5 and SSP5-8.5 scenarios were utilized.

The study's findings indicate a warming trend throughout Assam, coupled with increasing evapotranspiration demand and declining SPEI values, signifying heightened moisture stress during the rice-growing season. Yield variability is more significantly influenced by nighttime temperature extremes (TNX and TNN) and reductions in diurnal temperature range than by midday heat extremes. Multi-day extreme rainfall events (RX5day) negatively affect yields in flood-prone areas of the Upper and Middle Brahmaputra Basin and display mixed effects in regions with comparatively limited moisture; overall, precipitation extremes show substantial spatial variability. Scenario-based projections reveal greater yield volatility and an increased risk of yield decline under SSP5-8.5 compared to SSP2-4.5. This research framework provides a scalable and practical decision-support tool to enhance early warning systems for agro-meteorological variability, support climate-resilient agricultural planning, and inform evidence-based policy development.