Conference Agenda

Over the past two decades, high-throughput phenotyping (HTP) has transformed wheat research, enabling large-scale trait analysis and accelerating genetic gain. Yet its integration into official variety testing, the regulatory basis of varietal registration and seed certification, remains limited. Manual visual assessments still dominate, despite being labor-intensive, operator-dependent, and difficult to harmonize across sites and years. This gap hinders the operational deployment of digital phenotyping.

We report the validation and regulatory adoption of a handheld HTP system (Literal, HIPHEN) for automated wheat head counting under official field trial conditions. The system combines standardized RGB-NIR imaging with 3D reconstruction to deliver reproducible, plot-scale measurements. Validation under Bundessortenamt conditions covered seven trials, 79 plots, and 12 contrasting cultivars. Repeated image acquisition and reference manual counts were used to assess agreement, repeatability, and ranking stability.

Results showed strong agreement with manual measurements, high repeatability, and consistent cultivar ranking. Following validation, the system was formally integrated into official wheat phenotyping protocols, including the 2024 EPPO digital phenotyping guidelines, marking a shift from experimental tool to regulatory infrastructure and paving the way for scalable, standardized digital phenotyping.

Accurate and high-throughput assessment of photosynthetic capacity is a critical bottleneck in modern crop breeding. Traditional remote sensing methods often treat canopies as homogeneous surfaces, failing to capture complex 3D structural and physiological heterogeneity. This is particularly pronounced in rapeseed, where non-leaf organs like stems significantly affect light interception but contribute less to carbon assimilation. We propose a novel "structure-function" coupled framework to estimate canopy gross primary productivity (GPP). Using high-resolution UAV images, we reconstructed 3D point clouds and applied deep learning-based semantic segmentation to precisely separate leaf and stem points. A ray-tracing model calculated organ-specific light interception, while a vertical light use efficiency (LUE) profile, mapped from in-situ gas exchange measurements, assigned distinct physiological parameters to leaves and stems. Cumulative GPP was calculated by integrating diurnal dynamics over seven days. Results indicate that separating stems from leaves prevented GPP overestimation, as green stems intercept light but possess lower LUE. Furthermore, vertical LUE profiling corrected internal canopy shading errors. The model-derived biomass accumulation correlated strongly with destructively sampled ground-truth measurements (R^2 > 0.87). This demonstrates that combining 3D reconstruction with organ-level segmentation provides a robust proxy for photosynthetic capacity, enabling high-throughput phenotypic screening.

High-fidelity close-range spectral 3D reconstruction is pivotal for advanced plant monitoring, yet it faces persistent challenges regarding accurate radiometric calibration and the low spatial resolution of snapshot multispectral sensors. This study presents a multispectral reflectance field (MRF) framework for automatically generating radiometrically calibrated 3D point clouds of plants from snapshot multispectral imagery. We introduce a novel spectral ray model (SRM) designed to implicitly super-resolve imagery from filter-array cameras, effectively mitigating resolution constraints. Additionally, the framework integrates the segment anything model 2 (SAM2) to perform robust semantic segmentation, facilitating background exclusion and focused training. The automated calibration utilizes a hemispherical reference target to accurately decouple surface reflectance from the light field across varying normal vectors. Quantitative evaluations indicate high accuracy, with the calibration process achieving a spectral angle mapper (SAM) metric of 2.508° and a root mean square error (RMSE) of 0.048. Validation against standard ASD spectrometer measurements further confirms the system's precision, yielding a SAM of 5.093° and an RMSE of 0.033. By bridging the gap between raw snapshot data and 3D reflectance point cloud, this approach significantly enhances the reliability of 3D spectral analysis in agricultural research.

Existing three-dimensional (3D) plant point cloud segmentation methods struggle to accurately segment the complex structure of rice panicles. Achieving a balance between accuracy and real-time performance remains a challenge, which restricts the development of rice panicle phenotyping. To address these limitations, we propose a novel image-guided framework that leverages robust 2D instance segmentation to facilitate high-quality 3D instance segmentation, even on lower-quality datasets. Our method projects the 3D point cloud into 2D views, where a pre-trained instance segmentation model labels the projected points, generating sparse 3D seeds. The core innovation is an instance-aware diffusion process that propagates these seed labels to the full 3D cloud based on spatial, geometric, and feature similarities. A key motivation and advantage of our work is the ability to selectively use high-quality 2D images to guide the 3D segmentation process. The system can overcome incompleteness and noise inherent in poorer 3D datasets, enabling successful reconstruction where methods relying solely on 3D geometric information may fail. This work highlights the potential of bridging 2D semantic richness with 3D spatial understanding for scalable and robust 3D instance segmentation.

Microscopic phenotyping analysis is of great significance in the study of plant growth characteristics and environmental adaptability. However, traditional chemical staining has several limitations, including cumbersome preparation procedures, long processing time, reliance on chemical reagents, and damage to biological samples. In this study, an adaptive instance normalization module was innovatively introduced into an improved generative adversarial network framework. We explored the automatic transformation from unstained bright-field images or autofluorescence images to standard histochemical staining results. A paired dataset covering different tissues of wheat was constructed, and the model was trained to learn the feature distribution of tissue structures, thereby achieving cross-modal image translation.

The results demonstrate that the proposed virtual staining technique can accurately restore the key anatomical features of wheat tissues, and its visual effects and quantitative analysis results are highly consistent with those of physical staining. This method not only significantly shortens the experimental cycle and reduces chemical pollution, but also provides a novel digital and automated approach for the efficient acquisition of high-throughput plant microscopic phenotypes. It has important application value for accelerating wheat stress resistance screening and functional genomics research.

Reliable soil macronutrient maps are essential for precision agriculture and decision support systems (DSS). Yet models assessed with random cross-validation often overestimate accuracy due to spatial autocorrelation. In data-sparse regions, such bias can lead to unreliable fertilizer recommendations. We propose a stability-aware hyperparameter optimization framework for Digital Soil Mapping (DSM). Using CatBoost and Optuna, we implemented multi-objective optimization that jointly minimizes mean RMSE and its fold-to-fold standard deviation under strict spatial block cross-validation (5–25 km blocks, 10 km buffer). The approach was applied to map topsoil K, C, N, and P in Côte d’Ivoire. Conventional tuning produced unstable predictions once spatial independence was enforced. In contrast, Pareto-based stability optimization selected more conservative models (e.g., shallower trees, stronger L2 regularization), substantially reducing spatial variability with only minor loss in mean accuracy. Feature importance analysis showed that climate and topographic variables transfer more robustly across space than remote-sensing composites, which tend to overfit local signals. By explicitly quantifying the accuracy–stability trade-off, the framework acts as a safeguard against spatial overfitting and improves the reliability of soil maps for decision-making in data-scarce environments.