Agricultural theft is a serious issue in Japan, with traditional security measures like patrols and surveillance cameras facing challenges due to power constraints in farmland. Prior studies developed a Microphone-Based Surveillance System (MSS) that uses deep learning on compact, low-power edge devices to identify suspicious sounds. However, improving accuracy for a 3-class classification task to detect theft in farmland remained a challenge. This study addresses this challenge by employing Depthwise Separable Convolutions and Ensemble Distillation. The proposed model achieves an F1-score of 89.12% with a size of 48kB, demonstrating the feasibility of efficient AIoT devices for smart agricultural security.