This paper implements a district-scale multi-impact climate system for Karnataka, K-MIST (Karnataka Multi impact spatio-temporal framework) using CNN, ConvLSTM, Transformer, Fourier neural-operator, and graph-attention components. The archive contains 305 monthly 71 by 47 grids from January 2001 through May 2026 aligned with 31 districts. Models are trained until 2021, models are selected and calibrated on the non-overlapping period of 2022-2023, and tested on the untouched period of January 2024-May 2026. Hybrid's PR-AUC for drought was 0.129 and for flood was 0.365. For Flood, the gradient boosting algorithm achieved PR-AUC 0.371, while the hybrid algorithm performed better in terms of AUROC and macro-F1. May resulted in its own causal yield stage with a neural R-squared of 0.671, while the boosted ensemble achieved an R-squared of 0.674, MAE of 0.299 tonnes per hectare, and RMSE of 0.392 tonnes per hectare and RMSE 0.392 tonnes per hectare for the finalized 2021-2022 observations. Prospective 2023-2026 yields and 2027-2028 sensitivities from the climate models are separated from verified accuracy. The contribution is a reusable Karnataka workflow that reports evidence that is task dependent, calibrated with risk and uncertainty and with limits on labels without promising a universal winning model.

drought,food grain yield,convLSTM,transformer,fourier neural operator,graph attention,climate-impact assessment,Karnataka,flood,CMIP6