Multi-task geometric regression with agronomic priors for crop row and missing seedling detection in maize-soybean strip intercropping

Abstract

The strip intercropping of soybean and maize, characterized by planting the two crops alternately in adjacent rows, has been widely promoted in several regions of China due to its potential to enhance resource utilization efficiency and overall yield. Accurate detection of crop rows and missing seedlings is essential for enabling precision field operations such as variable fertilization and targeted spraying. Monocular vision has emerged as a core sensing modality owing to its low cost and high resolution. However, the significant differences in row and plant spacing between maize and soybean, coupled with complex field conditions such as weed interference and uneven emergence, severely limit the effectiveness of traditional image processing techniques based on thresholding and geometric fitting. These methods struggle to accommodate the morphological variability of multiple crops, resulting in poor row detection precision and unreliable identification of missing seedlings. In recent years, deep learning has shown strong performance in crop row detection and object recognition tasks, particularly through multi-task networks that integrate segmentation and localization-related features. Nevertheless, most existing studies focus on single-crop scenarios and often neglect the integration of agronomic knowledge, thereby limiting their robustness and interpretability in real-world field environments. To address these issues, this study proposes a Multi-Task Geometric Regression Network for row extraction and missing seedling detection in maize-soybean intercropping systems, built upon an improved U-Net++ architecture and guided by agronomic priors. The proposed method simultaneously performs crop segmentation, row direction prediction, and generation of a missing seedling heatmap. The geometric features output by the network are subsequently processed using agronomic prior-informed post-processing and geometric fitting to finally achieve row extraction and missing seedling localization. Agronomic constraints, such as row spacing regularity, are embedded in the loss function as prior-informed regularization terms, which further enhance detection accuracy and robustness in intercropped fields. Experimental results demonstrate that the semantic segmentation achieves an average Intersection over Union (IoU) of 0.82, an F1-score of 0.86, and a pixel accuracy of 0.91. Row centerline detection attains an F1-score of 0.86 and a mean offset (MO) of 3.9 pixels. For missing seedling detection, the crop classification accuracy reaches 0.91, the average localization error (ALE) is only 2.5 pixels, and the composite detection score (CD-F1) is 0.89. Compared with single-task methods without agronomic priors, the proposed multi-task framework exhibits significant improvements in both stability and accuracy for row detection and missing seedling localization in intercropping scenarios. These results provide practical guidance for deploying intelligent visual systems in precision agriculture and intercropping management.

Plant phenotyping relevance

作物画像から畝構造と欠株状態を抽出するマルチタスク手法を開発し、精度評価も行っており、植物状態の取得・推定が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.