Abstract Effective drought management is critical for agriculture‐based economies like India. This study examines whether a transformer‐based architecture can function as a robust pre‐trained backbone for operational drought forecasting across India. We utilized a pre‐trained transformer model, which was trained on the Indian Monsoon Data Assimilation and Analysis (IMDAA) precipitation data set (source domain) from 1979 to 2014. Transfer learning capabilities were evaluated by applying the pre‐trained model without target‐domain adaptation to 11 precipitation products (target data set) differing in source and spatial resolution. The target data sets were compared with the source domain to examine how well they reproduce regional climatic patterns. Transfer learning performance was evaluated over two distinct periods: the overall study period (2001–2019), partially overlapping with training, and the unseen testing period (2017–2019). Spatial evaluation is performed across six climatic zones delineated based on the Köppen‐Geiger classification. We utilized the 3‐month Standardized Precipitation Evapotranspiration Index (SPEI‐3) to quantify meteorological drought at a seasonal scale. The pre‐trained model maintained stable performance across data sets and climate zones, with correlations exceeding 0.7 during winter and pre‐monsoon seasons but declined during the monsoon. Transfer learning evaluation revealed consistent performance across most data sets, with ΔRMSE and ΔMAE within ±0.05 and ΔNSE within ±0.1 compared to the source domain. The composite skill score rankings showed that ERA5 and MSWEP were the closest to the IMDAA benchmark during the unseen period. The variation in performance across data sets underscores the importance of data set‐specific fine‐tuning for operational reliability.
Pathania et al. (Tue,) studied this question.