The research team, led by Nanda Kumar Tharu with Arsanchai Sukkuea and Pensiri Akkajit, assembled annual data from 1989 to 2024 covering 22 Terai districts. Climatic observations came from the Department of Hydrology and Meteorology, while crop statistics were sourced from the Ministry of Agriculture and Livestock Development and cross-validated against the National Statistics Office and Open Data Nepal. Eighteen climatic variables were considered, including precipitation, surface pressure, absolute and relative humidity, air and wet-bulb temperatures, maximum and minimum daily temperatures, land surface skin temperature, and wind speeds measured at 10 and 50 meters above ground.

A central methodological challenge was aligning climate data with crop biology. Rather than averaging weather over whole calendar years, the researchers weighted climatic variables according to crop-specific growing seasons: June to November for paddy, November to April for wheat, and March to September for maize. District-level values were aggregated to regional averages using area-weighted means based on the cultivated area of each crop, ensuring that agriculturally intensive districts contributed proportionally to the regional signal. Missing station-level gaps, which represented less than 2% of observations, were filled by linear interpolation, and annual yields were temporally disaggregated to monthly frequency using proportional allocation derived from crop growth stages.

Before estimation, the team confronted multicollinearity, the natural interdependence among temperature, humidity, pressure, and wind variables. An iterative Variance Inflation Factor (VIF) analysis eliminated redundant predictors, retaining only variables with VIF values below 10. Three separate crop-specific BVAR models were then estimated, each with seven endogenous variables and two lags, yielding a parameter-to-observation ratio of 0.42, well within statistical limits. The Minnesota prior was adopted to regularize the models, shrinking coefficients toward zero with stronger shrinkage applied to longer lags and cross-variable effects. Crucially, the prior’s hyperparameters were treated hierarchically, specified as Half-Cauchy distributions and estimated from the data rather than fixed arbitrarily.

Estimation relied on the No-U-Turn Sampler, an adaptive variant of Hamiltonian Monte Carlo, run across four Markov chains for 10,000 iterations each, producing 20,000 retained posterior samples. Convergence diagnostics were excellent: Gelman-Rubin statistics hovered near 1.0 for all parameters across all crops, effective sample sizes exceeded 400, and the Geweke diagnostic flagged only minor deviations in isolated cases. Sensitivity analyses varying the hyperparameters across plausible ranges confirmed that the ranking and direction of climate effects remained consistent, with coefficient variations below 5%. Alternative Cholesky orderings for the impulse response analysis, including crops-first and mixed arrangements, produced nearly identical results, with variations under 5%.

The impulse response functions revealed strikingly crop-specific sensitivities. Paddy yield responded most strongly to land surface skin temperature, declining by 0.456 t ha⁻¹ per one-standard-deviation shock, while maximum temperature and atmospheric pressure increased productivity by 0.441 and 0.418 t ha⁻¹ respectively. The authors interpret the positive temperature association cautiously, noting that moderate warming within the observed range can remain favorable for rice growth, consistent with field evidence from Tarahara, Nepal, where temperature increases of up to 2 °C benefited paddy before thresholds were exceeded. Negative associations with humidity and relative humidity align with known disease mechanisms, since humid microclimates favor pathogens such as Magnaporthe oryzae, the causal agent of rice blast.

Wheat emerged as the most wind-vulnerable crop. Wind speed variability at 50 meters showed a strong negative association with yield, and the forecast error variance decomposition identified wind-related variables as dominant drivers, with minimum wind speed at 50 meters contributing 28.4% and maximum wind speed at 10 meters contributing 22.1% of yield variability. The agronomic mechanisms are well documented: strong winds during heading and grain filling can cause lodging, accelerate evapotranspiration, and induce moisture stress even when soil water is adequate. Maize, by contrast, displayed moderate resilience, responding positively to maximum temperature and pressure but declining under excessive wind variability, high humidity, and elevated wet-bulb temperature, patterns consistent with heat stress effects on pollen viability and waterlogging constraints on root respiration.

Forecast performance varied by evaluation criterion. Maize achieved the highest point-forecast accuracy, with a mean absolute percentage error of 5.24% and RMSE of 0.214 t ha⁻¹, while paddy achieved the best probabilistic performance with the lowest log predictive score, indicating that its predictive distribution better represented observed outcomes and uncertainty. Benchmark comparisons showed the BVAR consistently outperformed classical VAR, ARIMA, and naive persistence models, with RMSE reductions of 13–16%, 9–13%, and 20–28% respectively. External validation against independent 2024 statistics from the Ministry of Agriculture and Livestock Development revealed percentage errors of −1.54% for paddy, −4.29% for wheat, and +3.58% for maize, all below 5%.

Perhaps the most consequential results concern the future. Under the low-emissions SSP1–2.6 scenario, all three cereals were projected to gain moderately by 2050, with increases of 7.4% for paddy, 9.4% for wheat, and 10.0% for maize. Under the intermediate SSP2–4.5 pathway, gains shrank, with wheat rising only 2.1%. Under the high-emissions SSP5–8.5 scenario, associated with a 3.2 °C temperature increase and 15% precipitation decline by 2050, yields were projected to fall by 11.9% for paddy, 18.0% for wheat, and 15.2% for maize. Wheat’s largest projected decline is consistent with its stronger historical associations with wind and temperature extremes, and the authors emphasize that these are scenario-conditional projections rather than guaranteed outcomes.

The study’s limitations are acknowledged by the authors: the model does not explicitly incorporate soil properties, irrigation infrastructure, fertilizer use, crop management, technological change, or pest dynamics, due to the lack of consistent long-term district-level data, and the estimated climate effects should be read as conditional associations rather than fully isolated causal effects. Chow tests detected significant structural shifts for wheat and maize around 2020, coinciding with the COVID-19 pandemic, though the authors note that supply-chain disruptions may be confounded with climatic and technological factors. Nevertheless, the framework demonstrates the value of hierarchical Bayesian inference for climate–yield forecasting in data-limited environments, and the authors argue that integrating such probabilistic forecasts into early-warning systems, seasonal climate services, and agricultural insurance could strengthen climate-risk management and national food security as the Terai’s agroecosystems face accelerating monsoon variability and thermal stress.

Subject of Research: Climate impacts on cereal crop production in Nepal's Terai region

Article Title: Modeling the impact of climatic variables on cereal crop production in the Terai region of Nepal

Article References: Tharu, N. K., Sukkuea, A., & Akkajit, P. (2026). Modeling the impact of climatic variables on cereal crop production in the Terai region of Nepal. Results in Engineering, 32, Article 113483. https://doi.org/10.1016/j.rineng.2026.113483

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113483

Keywords: Bayesian Vector Autoregression, Nepal Terai, cereal crops, climate variability, crop yield forecasting, paddy, wheat, maize, CMIP6 scenarios, agricultural adaptation, Modeling, impact