The data was generated using the Revised Universal Soil Loss Equation (RUSLE) from existing parameters data that influence soil erosion. The data focused on soil erosion of the Mellah watershed in northeastern Algeria for the years 1985, 1987, 1989, 1991, 1993, 1995, 1997, 2000, 2004, 2006, 2011, 2014, 2016, 2018, 2020, and 2022 to locate pattern spots of erosion. These pattern spots serve as informative inputs for a range of deep learning models—including LSTM, Transformer, CNN-LSTM, BiLSTM, GRU, CNN-only, Attention-LSTM, and ConvLSTM— aimed at enhancing soil erosion prediction accuracy.