Maize Tasseling Stage Automated Observation Method via Semantic Enhancement and Context Occlusion Learning

Accurately observing the maize tasseling stage (MTS) is critical for optimizing yield in agricultural ecosystem research. Despite some success in automated observation attempts, automating the tasseling stage remains an open problem. Existing methods lack robustness and fail to meet reliability requirements. To effectively address the heterogeneity in maize growth, this study proposes a new deep learning-based automatic observation method for the MTS, named TasselLFANet $^{\dagger } $ . It features a concise and efficient encoder-decoder structure, incorporating a novel receptive field synthesis (RFS) module to enhance the encoder’s responsiveness to semantic information, and a pair of parameter operators to improve task adaptability in the decoder’s upsampled features. Furthermore, the close intertwining of leaves and tassels during the tasseling stage results in frequent occlusion. We made an important observation that simulated occlusion learning with contextual information is highly effective. Therefore, we propose a clever image enhancement method called accurate-context occlusion, which is a simple copy-paste process but fully considers the contextual information in the image. To assess the effectiveness of the proposed method, we curated an open cross-domain MTS dataset, encompassing seven image series of continuous four-year sequences. Extensive experiments demonstrate that TasselLFANet $^{\dagger } $ achieves exceptional performance and outstanding accuracy in automatic tasseling stage observation, with a detection accuracy AP50 of 0.935 and a counting error MAE of 1.18, markedly surpassing advanced computer vision methods. This establishes it as an effective automated observation method for the MTS, providing a highly secure alternative to manual observations.