Flood risk has been increasing worldwide due to climate change and rapid urbanization. Rapid and accurate flood mapping is essential for reducing damage and supporting rescue activities. Synthetic Aperture Radar (SAR) has been widely used for flood monitoring because it can observe the Earth’s surface regardless of weather conditions or time of day. Conventional flood detection methods based on change detection between pre- and post-flood SAR images, however, often suffer from false detections caused by seasonal vegetation changes and speckle noise. This study proposes a flood detection method that generates a predicted SAR image representing the normal ground condition using a ConvLSTM model and compares it with an observed SAR image acquired during flooding. To preserve structural information while suppressing noise, the prediction model was trained using the Structural Similarity Index Measure (SSIM) as the loss function. In addition, a Siamese model was employed to model the correlation between predicted and observed images for flood change detection. Experimental results demonstrated that the proposed method reduced false detections and improved overall flood detection accuracy compared with conventional approaches. The results indicate that reducing the temporal gap between comparison images is effective for improving flood detection performance in SAR imagery.