DM_CorrMatch: a semi-supervised semantic segmentation framework for rapeseed flower coverage estimation using UAV imagery

Rapeseed (Brassica napus L.) inflorescence coverage is a crucial phenotypic parameter for assessing crop growth and estimating yield. Accurate crop cover assessment is typically performed using Unmanned Aerial Vehicles (UAVs) in combination with semantic segmentation methods. However, the irregular and variable morphology of rapeseed inflorescences presents significant challenges in segmentation. To address these challenges, advanced methods that can improve segmentation accuracy, particularly under limited data conditions, are needed. In this study, we propose a cost-effective and high-throughput approach using a semi-supervised learning framework, DM_CorrMatch. This method enhances input images through strong and weak data augmentation techniques, while leveraging the Denoising Diffusion Probabilistic Model (DDPM) to generate additional samples in data-scarce scenarios. We propose an automatic update strategy for labeled data to dilute the proportion of erroneous labels in manual segmentation. Furthermore, a novel network architecture, Mamba-Deeplabv3+, is proposed, combining the strengths of Mamba and Convolutional Neural Networks (CNNs) for both global and local feature extraction. This architecture effectively captures key inflorescence features, even under varying poses, while reducing the influence of complex backgrounds. The proposed method is validated on the Rapeseed Flower Segmentation Dataset (RFSD), which consists of 720 UAV images from the Yangluo experimental station of the Oil Crops Research Institute of the Chinese Academy of Agricultural Sciences (CAAS). The experimental results showed that our method outperforms four traditional segmentation methods and eleven deep learning methods, achieving an Intersection over Union (IoU) of 0.886, Precision of 0.942, and Recall of 0.940. The proposed semi-supervised learning-based method, combined with the Mamba-Deeplabv3+ architecture, demonstrates superior performance in accurately segmenting rapeseed inflorescences under challenging conditions. Our approach effectively handles complex backgrounds and various poses of inflorescences, providing a reliable tool for rapeseed flower cover estimation. This method can aid in the development of high-yield cultivars and improve crop monitoring through UAV-based technologies.