Evaluating maize emergence quality with multi-task YOLO11-Mamba and UAV-RGB remote sensing

• A novel multi-task framework for maize emergence quality estimation is proposed. • UAV-RGB dataset featuring maize seedlings at three early stages is constructed. • YOLO11-Mamba is proposed by introducing the cutting-edge Mamba attention. • OBB and KP are employed to accurately detect and pinpoint maize seedlings. • PC, ER, PS, and EU of maize emergence are analyzed and mapped in field-scale. Monitoring maize emergence quality is crucial for predicting yield, guiding precise management, and enhancing agricultural productivity and sustainability. Previous studies have highlighted the significant potential of UAV-RGB images and deep learning algorithms in evaluating maize plant counting (PC), plant spacing (PS), emergence rate (ER), and emergence uniformity (EU), respectively. However, a predominant limitation in these studies is their reliance on detected rectangular bounding boxes (RBB) and the associated geometric centers for quantifying emergence-related parameters. This approach tends to falter when confronted with the irregular growth orientations and diverse morphological attributes of maize plants, thereby compromising the accuracy and robustness of the estimations. To address this issue, this study builds a high-quality UAV-RGB dataset covering the V2 to V4 growth stages of maize seedlings and proposes a novel multi-task learning approach based on our YOLO11-Mamba framework. This methodology is committed to accurately detecting maize seedlings with oriented bounding boxes (OBB) and precisely pinpointing their centroids with key points (KP). Meanwhile, the Mamba Attention (MA) module, which combines a structured spatial model (SSM) and an independent spatial attention mechanism, is conducive to enhancing the state-of-the-art (SOTA) YOLO11′s adaptability to local features and global dependencies. The results indicate that multi-task YOLO11-Mamba outperforms YOLO11 in detecting maize seedlings, achieving a 2.1% improvement in OBB's AP 50 and a 2.9% promotion in KP's mAP. Furthermore, the R 2 for estimated PC, ER, PS, and EU is 0.915, 0.915, 0.908, and 0.754, respectively, demonstrating impressive performance in estimating maize emergence quality. Finally, a whole-field mapping is conducted to comprehensively demonstrate PC, ER, PS, and EU under different growth cycles based on multi-task YOLO11-Mamba. This study provides a smart and practical framework for the precise monitoring and management of maize in the early seedling stage. The dataset and code will be open-sourced at https://www.github.com/AG-WDS/MSQE . Evaluating Maize Emergence Quality with Multi-task YOLO11-Mamba and UAV-RGB Remote Sensing. (a) Characterization of maize emergence quality. (b) Data acquisition. (c) Data processing. (d) Multi-task YOLO11-Mamba. (e) Maize emergence evaluation on the test set. (f) Evaluation and mapping across entire field. (g) Potential applications.