Natural disasters pose an important threat toward human life and infrastructure because detection systems must be rapid and accurate. This paper classifies Cyclone, Earthquake, Flood, and Wildfire presenting a satellite imagery-utilizing deep learning-based disaster detection system. Synthetic images are produced using Conditional Generative Networks (CGANs). This overcomes dataset imbalance also betters model performance at once. Transfer learning fine-tunes such a ResNet-50 architecture that was trained on GAN-augmented datasets and original datasets. A wide-ranging evaluation then shows meaningful improvements with all of the augmented data as it used metrics. For improved model interpretability Grad-CAM visualization was integrated. It highlights image regions, also these regions influence predictions. Because it utilizes Flask, the best-performing model is deployed through a responsive web application, so users can upload images, and then the system is able to detect disasters in real-time. Deep learning makes automated disaster recognition scalable and explainable using this integrated system approach.