Conference Agenda

Accurate surface reflectance is fundamental for quantitative agricultural remote sensing, yet remains challenging under unstable illumination. This study proposes a real-time radiometric correction method integrating a UAV-mounted push-broom hyperspectral camera (400–1000 nm) with a synchronized upward-looking spectrometer (400–1300 nm). Cosine error correction and radiometric calibration were implemented to generate high-precision hyperspectral reflectance imagery under variable weather conditions.Beyond system validation, the corrected reflectance data were further applied to predict rapeseed (Brassica napus L.) pod maturity. Hyperspectral images acquired at different growth stages were analyzed using spectral feature extraction and machine learning models to classify immature, optimal maturity, and over-mature stages. Results showed that the proposed correction significantly improved spectral consistency across flight strips and enhanced model performance compared with traditional ELM-based imagery. The maturity classification achieved an overall accuracy above 90%, effectively distinguishing early-harvest (low oil content) and late-harvest (brittle pod loss) risks.The study demonstrates that high-precision hyperspectral reflectance imagery enables reliable maturity prediction in rapeseed, providing a practical decision-support tool for optimal harvest timing and yield maximization under field conditions.

Robotic harvesting in greenhouse environments is challenging because the fruit nodes (i.e. the cutting points) are often occluded by leaves and are difficult to perceive from a single camera view. Active-vision methods address this challenge by allowing robots to plan and move their camera such that the occluded nodes become visible. However, existing active-vision approaches treat all unseen regions as equally important and waste time exploring regions that are unlikely to contain fruit nodes. In this work, we tackle this limitation by proposing a prediction-guided active-vision method that uses prior knowledge about the plant structure and a few observed nodes to predict the 3D positions of the unobserved nodes and guides the camera towards these regions. On evaluations with a robotic arm with an RGB-D camera and eight tomato plants, the proposed method achieved an F1-score of 0.98 in predicting the node positions and perceived 89.2% of the fruit nodes within three views, which was 7.4 percentage points more than an active vision planner without structural priors and 22.9 percentage points more than a random planner. These results demonstrate that using structural priors with active vision can focus search and speed up the process of robotic perception for harvesting.

Detailed digital twins of fruit trees acquired at different growing stages are highly valuable for spray simulations and evaluating orchard management strategies. Therefore, the aim of this study was to develop and evaluate a LiDAR-based measurement and analysis pipeline for multi-temporal 3D reconstruction and phenotyping of orchard trees in commercial high-density systems. A custom platform combining a LiDAR sensor with RTK-GNSS was used to acquire multi-view scans of 22 pear (cv. Conference) trees over two years in Sint-Truiden, Belgium. Measurements covered three common training systems, namely V-hedge, three-branched chandelier, and free spindle, under winter, spring, and summer (post-harvest) conditions. The proposed pipeline registers and merges multi-view point clouds into tree-level representations and generates watertight meshes using a shrink-wrapping approach to obtain a fine-grained representation of canopy and branch structure. The resulting point clouds and meshes are further analysed to extract branch skeleton structure and canopy segmentation, while also providing a basis for additional parameter estimation such as leaf area density. The obtained results show that the workflow can accurately reconstruct the tree geometry across growing systems and seasonal stages despite varying canopy conditions. These findings support the use of LiDAR-based digital twins for tree-level phenotyping and detailed spray simulations.

Abstract. Smart agriculture technologies offer innovative solutions for biomass monitoring in Mediterranean perennial crops characterized by high spatial variability and climate constraints. This study aimed to estimate above-ground bio-mass of medlar (Mespilus germanica L.) orchards in the Mediterranean region using Unmanned Aerial Vehicles (UAVs) equipped with multispectral sensors. UAV flights were conducted during the peak vegetative stage to acquire high-resolution imagery. Vegetation indices (NDVI, NDRE, and GNDVI) and canopy structural parameters derived from photogrammetric point clouds were extracted and correlated with field-measured biometric variables, including canopy volume and destructive biomass sampling. Regression models and multivariate analysis were applied to assess the predictive capability of spectral and geometric features. Results showed strong correlations between UAV-derived parameters and meas-ured biomass (R² > 0.80), with combined spectral–structural models providing the highest estimation accuracy. The approach allowed site-specific biomass mapping, highlighting intra-orchard variability linked to soil and microclimatic conditions typical of Mediterranean environments. The findings demonstrate that UAV-based remote sensing represents a reliable, non-destructive, and scalable tool for biomass estimation in minor fruit species, supporting precision orchard management and sustainable resource use.

Foreign objects such as soil clods and stones generated during mechanized garlic harvesting reduce product quality and increase post-harvest processing costs. Collection-type garlic harvesters require two to four workers to manually remove foreign objects, resulting in high labor costs and limited efficiency. In this study, to develop an image recognition–based automatic foreign object detection system applicable to a collection-type garlic harvester, a recognition model was designed using a CNN-based commercial deep learning tool and evaluated. To ensure applicability in outdoor environments, images were collected under natural lighting(8,969–48,150 lx) using an RGB camera. To simulate mechanical harvesting conditions, garlic bulbs and foreign objects were placed on the harvester conveyor, and their transport was captured. A total of 3,000 images were acquired and annotated using bounding boxes. The dataset was divided into training(70%) and validation(30%) sets. The model achieved an F1-score of 0.917 with a training loss of 0.043. The false negative and false positive rates were 2.64% and 2.82%, respectively, with an average inference time of 92.32 ms. These results demonstrate the feasibility of foreign object detection under open-field conditions. Further research will integrate the detection system with a mechanical removal device for automated sorting.