This study, conducted by the Inner Mongolia Alpine Crops Research Center, focuses on the development of efficient and precise soybean-corn intercropping technology alongside its supporting equipment. An STM32 device (powered by ARM architecture) was employed as the core hardware platform, running the Ubuntu operating system to support a lightweight visual recognition system. The Transformer deep learning model was trained on independently collected soybean and corn leaf datasets and was deployed efficiently on the device. Experimental results demonstrate that the model achieved a classification accuracy of 100% for soybean and corn identification tasks under standard conditions, and maintained over 98% accuracy in various challenging environmental scenarios, including low-light and complex background conditions. Additionally, the system exhibits excellent generalization ability and robust performance in multi-crop classification tasks, achieving an accuracy of 96.5% when extended to five crop categories. The real-time recognition and decision-making process on the embedded platform ensures timely and accurate pesticide application with minimal latency. This research provides a technological foundation for the precise management of soybean-corn intercropping in cold regions and highlights the potential of embedded devices in advancing intelligent agricultural practices.