Precision management in high-density orchards requires individual-tree, nondestructive monitoring of canopy nitrogen concentration (CNC), but hyperspectral applications are limited by two factors: unmodeled vertical stratification of CNC within 3D canopies and mixed-pixel effects near canopy boundaries. We develop a cross-modal framework that co-registers RGB-derived 3D point clouds with hyperspectral orthomosaics, enabling individual-tree localization in dense orchards. With this framework, we quantified layer-specific nitrogen-spectral relationships and assessed mixed-pixel effects across canopy positions. Stratified sampling, continuous wavelet transform (CWT), and partial least squares regression (PLSR) with variable importance in projection (VIP)–based band selection were used for spectral optimization, and K-means was applied to isolate representative canopy pixels. Field experiments over two consecutive years (2023–2024) revealed consistent CNC gradients, with the lower canopy exceeding the upper by 0.5–9.5 % across fertilization treatments. CWT-2 delivered the most accurate and robust performance across years. VIP-PLSR indicated layer-dependent CNC-informative wavelengths spanning the visible, red-edge, and near-infrared regions, with scale-dependent cross-layer overlap after CWT. Pixel clustering revealed distinct spatial structure: canopy-interior pixels exhibited characteristic vegetation spectra and achieved R_val² of 0.69–0.76, substantially outperforming boundary-affected pixels with R_val² of 0.48–0.57. These results demonstrate that coupling spectral feature optimization with layer-specific modeling and clustering-based pixel screening improves the accuracy of tree-level CNC estimation in complex canopies. The proposed framework provides a mechanistic and operational basis for robust biochemical retrieval in structurally complex orchard systems.

Noninvasive analysis of pod phenotypic traits under field conditions is crucial for soybean breeding research. However, previous pod phenotyping studies focused on postharvest materials or were limited to indoor scenarios, failing to generalize to real-field environments. To address these issues, this paper employs an instance segmentation approach for the precise extraction of the pod area from multiplant RGB images in preharvest soybean fields. We first introduce a cost-effective workflow for constructing datasets of densely planted crop images with a uniform backdrop. Starting with video recording, high-quality static frames are collected by automatic selection. Then, a large vision model is explored to facilitate dense annotation and build a large-scale soybean dataset comprising 20k pod masks. Second, the pod instance segmentation model PodNet is developed based on the YOLOv8 architecture. We propose a novel hierarchical prototype aggregation strategy to fuse multiscale semantic features and a U-EMA prototype generation network to improve the model's perception performance for small objects. Comprehensive experiments suggest that lightweight PodNet achieves a superior mean average accuracy of 0.786 in the custom pod segmentation dataset. PodNet also performs competitively on in-field images without a backdrop and enables real-time inference on the edge computing platform. To the best of our knowledge, PodNet is the first pod instance segmentation model for preharvest fields. The low-cost and high-precision extraction of pods is not only a prerequisite for phenotypic analysis of the pod organs but also constitutes an important foundation in conducting cross-scale phenotyping from whole-plant to seed levels.