Abstarct: Drought intensifies the water crisis and causes irreparable damage to communities. In recent years, machine learning methods have been considered by researchers for drought assessment. The goal of this study is drought assessment in Zabol and Zahedan cities during (1990-2020). In this regard, standardized precipitation index (SPI) at seasonal and annual time scales, partial autocorrelation analysis (PACF), and random forest algorithm (RF) were employed. After SPI calculation, the results of PACF analysis for SPI are used as inputs of the model. Training and testing data were verified with two different inputs. According to the results on a seasonal and annual scale of SPI, the normal drought category (N) has almost the highest frequency in both stations and drought intensity in Zahedan station was higher on seasonal and annual scale than Zabol station and the peak of annual drought in both stations occurred in 2001. Also, the peak of drought in both stations on seasonal scale occurred in seasons without rainfall. According to the PACF analysis, the study areas have undergone climate change over 30 years. The results of the model were evaluated by several statistical parameters. The index of agreement (IOA) of Zabol station for training data by considering four time-lags (2, 4, 6, and 12 months) and three time-lags (2, 4, and 6 months) as input, were 0.9648 and 0.9256, respectively and for testing data was approximately 0.8556 and 0.8673, respectively. IOA at Zahedan station for training data with four time-lags (2, 4, 6, and 8 months) and three time-lags (2, 4, and 6 months) were 0.9495 and 0.9205, respectively, and for testing data were 0.7408 and 0.6303, respectively. Other statistical parameters also indicate the permissible accuracy of the RF model in SPI estimation.