Flood monitoring and early warning systems (FMWS) are very vital for reducing the effects of natural catastrophes. This paper offers a sophisticated IoT-FMWS-TARGCNN-AG method combining Graph Convolutional Networks (GCNs) and Temporal Attention-based Recurrent Neural Networks (RNNs) for improved flood forecasting. The suggested solution employs IoT sensors coupled to a NodeMCU for real-time data collecting and low-latency transfer. While GCN catches spatial relationships, limiting false alarms, the RNN Temporal Attention technique reduces processing delays by prioritizing relevant information. Experimental findings reveal that IoT-FMWS-TARGCNN-AG achieves up to 28.96% reduced latency, 30.78% greater accuracy in flood prediction, 28.78% lower false alarm rate, and 30.58% enhanced packet delivery ratio compared to current approaches such as IoT-RFT-PS, FF-ML-IoT, and LoRaWAN-IoT-FMWS. Additionally, the Receiver Operating Characteristic (ROC) study indicates a 25.36% gain in system adaptability over rival models. These findings demonstrate the usefulness of the proposed model in delivering highly accurate, low-latency, and dependable flood prediction and alerting, making it a viable tool for real-time disaster management applications.