Abstract: To address the challenges of low efficiency, limited spatial coverage, and unstable data accuracy in manual tobacco planting area surveys, this paper proposed a novel method for large-scale tobacco plant identification and counting that integrates remote sensing imagery with intelligent algorithms. Firstly, the Segment Anything Model 2 (SAM2) was introduced for weakly supervised field segmentation and accurate extraction of field parcel boundaries, addressing the issues of blurred field boundaries and regional redundancy in remote sensing imagery. Secondly, A YOLOv11-based module of stage recognition was used to screen images of growth stages, identifying and selecting field parcel regions with distinct tobacco plant morphological features, achieving an accuracy of 99.4%. Finally, in the YOLOv12-based tobacco plant identification and counting stage, an image splitting strategy was employed for dense object detection. Through region fusion and NMS deduplication, the recognition accuracy and efficiency were further improved. The method exhibited excellent performance in large-scale tobacco plant identification and enumeration, achieving a mAP@0.5 of 91.5% and a Precision of 97%, which validated its effectiveness in improving recognition accuracy and efficiency. The proposed method for tobacco plant identification and counting, which integrates remote sensing imagery and intelligent algorithms, can effectively enhance the accuracy and efficiency of tobacco plant identification in large-scale tobacco field scenarios. It features excellent spatiotemporal scalability and technical application value, providing reliable data and algorithmic support for tobacco cultivation supervision.