The facilitate sustainable and efficient urban waste collection in smart city environments, the real-time monitoring of bins, prediction of bin overflow, adaptation of routes and intelligent decision execution are essential for smart waste management systems based on IoT devices. But the current approach of Deep Learning (DL) classification visual detection and static forecasting are limited by the functional modules, lack of multi-sensor fusion and have no closed-loop decision-making, which makes it difficult to be extended to complex scenes under dynamic urban waste. To address these shortcomings, this research introduces the DeepSense Hierarchical Smart Bin Monitoring Network (DHS-BMN) that use heterogeneous IoT sensing, hybrid CNN-LSTM-Attention overflow prediction, Reinforcement Learning (RL) based on adaptive routing and region-aware clustering of bins based on the K-Means algorithm, in a unified closed-loop framework for intelligent urban waste management. The results of the experiments verified the suitability, robustness and scalability of the proposed DHS-BMN for the next generation smart city waste management deployments, as it obtained overall higher accuracy in the overflow prediction (97.8%) and efficiency in routing (95.1%) when compared with the existed method.