DiffKNet-TL: Maize phenology monitoring with confidence-aware constrained diffusion model based on UAV platform

Abstract

Agricultural phenological monitoring is key to understanding climate change impacts on ecosystems and optimizing crop management. UAV remote sensing provides high-resolution phenotypic data at the landscape scale, resolving plant-level phenological characteristics. However, weather and operational constraints hinder high-frequency UAV imaging on a daily scale. Previous studies have often focused on single phenotypic traits at single growth stages. Therefore, dynamic changes of multiple traits throughout the entire growth cycle were neglected, resulting in the identification of only partial phenological events. This study proposed a DiffKNet-TL (Diffusion-enhanced K-Net for tassels and leaves) model to identify maize phenology throughout the entire growth period, trained solely on tasseling stage images. A maize dataset covering leaves and tassels from seedling to harvest was also constructed. To capture dynamic phenotypic changes across scales and time, the K-Net architecture was also enhanced with a Transformer-based backbone and a dynamic loss reweighting strategy. For small size tassel targets, a confidence-aware discrete diffusion module was integrated, refining contours and reducing artifacts. Results showed that the implement of Swin-Transformer improved mIoU by 4.75%, and SSLossIoU added another 1.82%. Tassel refinement brought a further 2.55% IoU gain. Final IoUs for background, leaves, and tassels reached 84.98%, 75.85%, and 65.77%, respectively. DiffKNet-TL also generalized well across growth stages and remains robust under occlusion, straw residue, lighting disturbances, and uneven yellow leaf coloration. This study can provide the technical foundation and data support for large-scale, automated phenological monitoring of maize leaves and tassels, aiding in the quantitative dynamic tracking of maize phenology.