Drought conditions in the Mekong River Basin (MRB) have been exacerbated by increasing hydroclimatic variability, posing significant challenges to water resources, agricultural productivity, and ecosystem stability. Although the application of deep learning techniques in drought forecasting is on the rise, a thorough understanding of the interactions among meteorological, hydrological, and agricultural drought processes across various temporal scales remains limited, particularly in the MRB. This study introduces a comprehensive multi-index drought forecasting framework that integrates causal inference, statistical trend analysis, explainable artificial intelligence, and a comparative analysis of eight deep learning architectures. It examines three standardised drought indices, including the Standardised Precipitation Index (SPI), Standardised Runoff Index (SRI), and Standardised Soil Water Index (SSWI), across short-, mid-, and long-term accumulation scales at six mainstream hydrological stations from 1982 to 2019. Forecasting experiments were conducted to assess one- and two-year prediction horizons under consistent training conditions. The results of model benchmarking indicated that forecasting performance was more influenced by drought type and station characteristics than by the prediction horizon or lookback duration. Notably, the PatchTST, RNN and KAN models consistently demonstrated the strongest performance across multiple drought indices and timescales. The explainability analysis highlighted that the models exhibiting high performance attributed outcomes primarily to early antecedent conditions and specific temporal intervals linked to drought persistence. These findings underscore that effective drought forecasting in the MRB fundamentally relies on the interplay between drought memory, hydrological variability, and the ability of forecasting architectures to capture long-range temporal dependencies.