Reliable wireless link estimation is crucial for ensuring robust communication in resource-constrained Internet of Things (IoT) environments. However, traditional approaches often struggle with precision, computational efficiency, and adaptability in the dynamic conditions of IoT networks. Additionally, existing wireless link datasets suffer from several limitations, including fragmentation across multiple sources, class imbalance, and limited category diversity. In this paper, we propose a lightweight deep learning model based on the CoAtNet architecture, which integrates depthwise convolutions and attention mechanisms to enhance both efficiency and accuracy in link quality estimation. To provide a centralized and standard data resource for wireless link estimation research, we develop a data loader and introduce a data collection methodology that generates a more granular and balanced dataset. Experimental results across five datasets demonstrate the robustness and generalizability of our approach. Notably, our model achieves fine-grained precision and high efficiency on the Raspberry Pi 4, attaining an accuracy of 99.22% while using only 25% of RAM and 36% of CPU resources.