This paper proposes a ConvLSTM neural network architecture to forecast precipitation over the next 6 hours. Satellite data and terrain information are fed as inputs to capture atmospheric characteristics over time and space. The model is first trained generally then retrained specifically for a domain in Brazil to improve forecasts. Evaluation of forecasts for December 2022 showed the retrained model had higher precision, recall, and F1-scores, suggesting targeted training can enhance localized predictions. Future work will focus on increasing general training and applying the approach to other forecasting domains and data sources.