Monthly rainfall in tropical regions exhibits complex spatiotemporal variability, requiring models to simultaneously capture spatial heterogeneity and temporal patterns. This study compares ConvLSTM, ConvLSTM with Temporal Embedding (ConvLSTM-TE), and 3D CNN for monthly rainfall prediction using CHIRPS v2.0 data (1990–2025) over a 20 × 20 grid domain (5.75°–6.75° S; 106.5°–107.5° E) covering Bekasi and its surrounding areas. Twelve monthly rainfall maps were used to predict the subsequent month's rainfall map, with consistent temporal splits, forecast horizons, and evaluation metrics across all models. Hyperparameter tuning covered the number of filters, learning rate, batch size, scaling methods, and recurrent dropout, the latter being tested only for recurrent models. ConvLSTM-TE incorporates a 16-dimensional calendar representation derived from month and season embeddings, along with sine–cosine features. Climatology and persistence were used as baseline benchmarks. Across the 400-grid domain, the best ConvLSTM-TE configuration, selected based on test-set RMSE, achieved an RMSE of 82.368 mm/month, an MAE of 59.024 mm/month, a correlation coefficient (r) of 0.868, an R² of 75.25%, and an RSR of 0.497. It was followed by ConvLSTM (RMSE: 86.048 mm/month) and 3D CNN (RMSE: 96.347 mm/month). All three models outperformed persistence (RMSE: 125.639 mm/month), but only the two ConvLSTM-based models outperformed climatology (RMSE: 88.336 mm/month), achieving RMSE skill scores of 6.8% for ConvLSTM-TE and 2.6% for ConvLSTM. ConvLSTM-TE also achieved the lowest median RMSE across 92 training runs and retained the lowest RMSE under validation-loss-based model selection, although its performance advantage over ConvLSTM was marginal. The models preserved spatial gradients and seasonal phases but tended to smooth extreme rainfall events. An independent evaluation of ConvLSTM-TE against CHIRPS data for January–June 2026 yielded a mean absolute error of 57.15 mm/month, comparable to its test-set performance. The selected model was implemented in a Streamlit prototype for visualization.