PPNet: a deep learning framework for rice seedling localization and distribution from UAV orthomosaics

Field-scale information on rice seedling distribution is crucial for transplanting quality assessment, site-specific seedling-stage management, and precision agricultural operations. However, seedlings in UAV orthomosaics are small, closely spaced, and easily confused with soil, shallow water, plastic mulch, and shadows, which makes accurate localization challenging. Meanwhile, under high-density planting, existing segmentation-based networks often generate adhesive responses between neighboring seedlings, leading to missed detections and biased counting. To address these issues, this study proposed PPNet, a deep learning framework for rice seedling localization and distribution mapping from UAV orthomosaics. The framework uses root-based point annotation and integrates a rice-point perception module (RPPM), an adjacent decoupling module (ADM), and an aggregated attention module (AAM). These modules enhance seedling-point responses, separate adjacent seedlings, and introduce row-structure information. Experiments were conducted on a UAV orthomosaic dataset containing 3,013 image tiles and 220,553 seedling point annotations from different rice-growing regions. PPNet achieved an F1 score of 0.9029, an AP of 0.8881, a counting MAE of 6.99, and an R2 of 0.9553. The method also generated field-scale seedling point maps and density maps under different planting densities and field backgrounds. The results demonstrate that PPNet can provide reliable seedling localization and counting information for crop establishment assessment and targeted seedling management, supporting intelligent agricultural machinery decision-making in precision farming operations.