A Segment Anything Model Approach for Rice Seedlings Detection Based on UAV Images

Abstract. Accurate estimation of regional rice yields is crucial for food security and efficient agricultural management. In this regard, the use of Unmanned Aerial Vehicles (UAVs) that have revolutionized crop monitoring by providing high-resolution images for precision agriculture, is beneficial. This study explores the potential of Segment Anything Model (SAM) for detecting rice seedlings, focusing on determining the optimal approach and prompt for this task. We examined three SAM scenarios: automatic mask generation, bounding box prompt, and point prompt. Our evaluation criteria included processing time, visual interpretation, and accuracy indexes. The results demonstrated the effectiveness of SAM in rice seedling detection, highlighting the importance of selecting the appropriate prompt for specific agricultural applications. Our findings reveal that the point prompt method emerges as the preferred choice for rice seedling detection, offering superior accuracy and reliability. Specifically, it achieved mIoU and mDice scores of 94.57 % and 0.97, respectively. While the bounding box approach showed promise, despite slightly lower precision, it may still be suitable depending on application-specific requirements. Conversely, the automatic mask generation scenario proved unsuitable for this task due to its low accuracy and inability to effectively detect rice seedlings. The outcomes of this study serve as a baseline for evaluating SAM prompts, guiding future improvements and refinements to enhance its performance in real-world agricultural applications.