MSNet: A Novel Deep Learning Framework for Efficient Missing Seedling Detection in Maize Fields

Machine vision application in agriculture has spurred significant interest in crop-missing detection. This study targets critical challenges, such as comprehensive aerial imagery coverage, tiny seedlings easily mistaken for weeds, and the absence of adaptive learning in traditional row classification. We propose a novel methodology for missing seedling detection, which comprises three essential stages: seedling localization, row classification, and missing region prediction. We present SeedNet, a detector that leverages row direction information to enhance small seedling detection performance. By incorporating this information, SeedNet improves recall by 25.3% and AP by 15.4% compared to the baseline. Additionally, we introduce PeakNet, a deep learning-based classifier for row segmentation that efficiently adapts to row spacing without any prior assumptions, achieving an accuracy of 99.69%. Finally, missing areas are predicted by analyzing the relative distances between adjacent seedlings within the same row. Under challenging outdoor conditions, this method achieves a missing detection accuracy of 98%, meeting practical requirements for field testing. SeedNet and PeakNet demonstrate exceptional performance in real-time processing, achieving inference speeds of 105 FPS and 2295 FPS, respectively. These results indicate strong potential for practical applications in real-time agricultural systems. The proposed approach provides a high-performance, low-cost solution for crop missing detection.