Comparative Performance of ContextCNN and Transformer Models for Flood Inundation Extent Prediction Using Sentinel-1, MODIS, and GPM Data in the Jabodetabek Metropolitan Area, Indonesia

Floods are a recurring, structural hazard in the Jabodetabek metropolitan area, driven by land subsidence, rapid land-use conversion, poor drainage, and increasingly intense rainfall. Conventional monitoring relies on sparse, point-based gauges and warnings that are issued only shortly before overflow. This study develops and compares two deep learning architectures, a Contextual Convolutional Neural Network (ContextCNN) and a Transformer, to predict flood inundation extent by integrating Sentinel-1 SAR backscatter, MODIS vegetation and water indices, and GPM IMERG rainfall accumulations. The workflow follows CRISP-DM. Water labels were derived from Sentinel-1 VV backscatter with a −15 dB threshold, and 137 scenes from January 2023 to June 2026 were split chronologically into training, validation, and test periods. Both models were benchmarked against a persistence baseline and Random Forest at pixel and grid-cell (km²) levels. On the test set, the Transformer and ContextCNN achieved PR-AUC values of 0.8424 and 0.8401, outperforming Random Forest (0.8283) and persistence (0.4169). For newly inundated pixels, where persistence fails completely, the two models reached PR-AUC values of 0.7310 and 0.7125. At grid level, ContextCNN produced the lowest MAE (0.0163 km²) and MAPE (38.64%), while the Transformer achieved the lowest RMSE (0.0740 km²) and highest R² (0.8602). Because predictive quality was practically equivalent while ContextCNN trained 3.6 times and inferred 5.9 times faster, ContextCNN is the more practical option and was deployed as a prototype web application.