Bayesian assessment of return levels for observed temperature extremes across Odisha in India

Temperature-extreme risk is often assessed through trend-based indices, but regional climate-risk planning also requires uncertainty-aware estimates of rare thermal events and their spatial variability. This study presents a grid-wise Bayesian extreme-value assessment of observed temperature extremes over Odisha, India, using IMD daily maximum and minimum temperature data for 1951–2024. Annual and seasonal hot-day, warm-night, and cold-night block-extreme series were constructed for retained 1° × 1° grid cells. Stationary and nonstationary generalized extreme value models were fitted independently to each grid-series combination, with standardized time, ENSO, and IOD covariates evaluated as predictors of the location parameter. Model selection prioritized diagnostically clean candidates, retaining explicit cautions where needed; posterior 10-, 20-, and 50-year return levels were estimated for early and recent climate states. Of 30 Odisha-overlap grids, 28 were model-ready (93.3% spatial model-ready support); all 140 selected grid-series fits yielded numerically valid return levels. Annual and March–June daytime maxima were nearly identical, confirming concentration of annual hot-day extremes within the extended pre-monsoon heat season. The recent spatial median 50-year TXx MAMJ return level was 43.81 °C; selected stationary M0 models impose zero early-to-recent change. Warm-night extremes showed stronger time-varying model support, and recent TNx JJAS return levels were lower than early-period values. Common-M1 sensitivity indicated small positive but generally uncertain TXx changes, coherent negative TNx changes among eligible fits, and near-zero/mixed TNn changes. Within diagnostically eligible matched support, negative TNx coefficients persisted in the 1951–2007 sensitivity (12/12 Annual; 19/19 JJAS grids). Teleconnection-supported models were sparse and showed mixed coefficient evidence. Return levels provide conditional fitted-tail summaries of observed extremes, not deterministic forecasts.