This study focuses on nowcasting cloud-to-ground lightning in the Amazon region, utilizing the Green Ocean Amazon (GoAmazon) campaign data. Lightning strikes can have severe socioeconomic consequences, making accurate forecasting systems crucial for risk assessment and decision making. Deep learning models have shown promising results in severe storm forecasting and are widely used in various meteorological applications. Overcoming the challenges posed by the chaotic and highly variable nature of lightning events, different deep learning models were explored for prediction. The performance of the models was compared on the basis of the architecture and data configurations, and the evaluation methods considered spatial variability and sensor accuracy. The study sought to develop a very short-term (up to 12 minutes) forecasting system for cloud-to-ground lightning in the Amazon Basin, taking into account the distinctive characteristics of the event and the uncertainties associated with predicting lightning strikes. Considering various probability thresholds for the occurrence of lightning strikes, the best thresholds were determined for two artificial neural networks, U-Net and ConvLSTM. The validation methods used for model evaluation were also important in this study, given that these methods make validations more assertive due to the consideration of characteristics associated with lightning (e.g., spatial variability) and sensors (e.g., spatial accuracy). Both models tracked lightning effectively, with the best model based on U-Net achieving 0.642 (POD) and 0.353 (FAR), and the best model based on ConvLSTM reaching 0.689 (POD) and 0.307 (FAR). Also, the smallest spatial error (e.g. < 5 km) based on the overlap analysis was also observed for these thresholds. Therefore, it is concluded that the models tested are capable of predicting lighting in the short term in the Amazon region.