Unsupervised geometry-driven DSM terrain correction for Rice seedling segmentation in flooded Paddy fields

Rice seedlings in flooded paddy fields are small with low ground cover. Specular reflections and weakened texture features substantially compromise their distinguishability in remote sensing imagery. Furthermore, digital surface models (DSMs) generated from unmanned aerial vehicle (UAV) imagery often contain systematic trend errors, which hinder reliable seedling segmentation. To address these challenges, we propose an unsupervised, geometry-driven method for segmenting rice seedlings from UAV imagery. This method formulates rice seedling segmentation as a geometry-driven seedling structure segmentation problem. An uncertainty-weighted DSM terrain-structure decomposition model is constructed for this purpose. By incorporating second-order total variation constraints, it corrects low-frequency terrain trends, thereby separating ground trends from sparse structures such as seedlings under unsupervised conditions. Moreover, the corrected geometric information is fused with morphological cues from RGB imagery to achieve stable segmentation of rice seedlings. This approach eliminates the need for ground control points or pixel-level manual annotations and effectively suppresses interference from water surface reflections and complex backgrounds. Experiments across multiple typical paddy field scenarios show that our method achieves seedling segmentation accuracy comparable to that of state-of-the-art fully supervised methods under complex conditions, attaining an mIoU of approximately 0.85. In addition, it exhibits stable structural consistency and robustness, validating its feasibility for large-scale paddy field monitoring and autonomous agricultural machinery navigation applications.