Abstract In order to improve the reliable monitoring and prediction of drought behavior, it is crucial to comprehensively consider composite indices. Initially, four univariate drought indices were evaluated: the Standardized Precipitation Index, the Standardized Soil Moisture Index of the top two layers (SSI1 and SSI2), and the Standardized Precipitation Evapotranspiration Index. Subsequently, this paper introduces five composite drought indices: the Multivariate Standardized Drought Index, the modified Aggregate Drought Index, the Joint Drought Deficit Index, the Machine Learning (ML-based) Drought Index (SVMs), and the Artificial Intelligence (AI-based) Drought Index (ANF-PSOs) models based on precipitation (P), soil moisture at two layers (SM1 and SM2), and reference evapotranspiration as input variables. These drought indices are formulated using P-ETref inputs, P-SM1 inputs, and P-SM2 inputs. The study covers 30 main basins classified into six climate zones, including coastal wet, mountain, semi mountain, semi desert, desert, and coastal desert across Iran from 1979 to 2021. The performance of the studied models was evaluated using the correlation coefficient and root mean square error in comparison with SPI at the same time scales as the target model. Drought characteristics, including the number of drought events, duration, frequency, and intensity, were determined for each model at monthly, seasonal, and yearly time scales. The results revealed that the recommended P-SM1 inputs for SVMs and ANF-PSOs models significantly outperformed the MSDIs, modified ADIs, and JDIs models. The results indicated that the introduced composite models effectively captured the comprehensive SM situation without being heavily influenced by individual parameters. The behavioral patterns of these indices remained consistent, except for the specific performance of ETref, which caused some inconsistencies. Moreover, a comparison among different climates revealed that ETref played a prominent role in the discrepancies observed in the output of the composite models, resulting in a strong relationship between P and ETref. Consequently, when constructing composite indices, the information conveyed by ETref was more readily disregarded by JDI and ADI but retained in ANF-PSO. This research sheds light on the mechanisms of these ML-based and AI-based composite approaches in integrating different drought features. It offers valuable insights into the performance of composite drought models and provides benchmarks, particularly in dry climates, to enhance drought monitoring methods.
