High-resolution multi-sensor UAS framework for individual tree health monitoring and structural analysis in walnut orchards

Persian or English Walnut (Juglans regia L.) growing in Hungary faces significant challenges from complex biotic pathogens and abiotic climate stressors. While farmers possess the expertise to identify these pathologies, early diagnosis is often impeded by the physical inaccessibility of the upper canopy, where symptoms frequently manifest first. To overcome these limitations, this study proposes a multi-sensor Unmanned Aerial System (UAS) framework capable of acquiring high-resolution geospatial data to identify physiological features invisible to the human eye.

The research was conducted at two distinct sites in Central Hungary, representing contrasting management regimes. The first is a 2.4-hectare intensive commercial orchard utilizing rigorous irrigation and chemical protection. The second is a 4.2-hectare genetic archive owned by the HUALS Fruit Growing Research Center; this site contains diverse cultivars with varying management levels (including untreated controls), offering a higher probability of observing heterogeneous disease responses.

Data acquisition utilized a DJI Matrice M210 equipped with a 10-band MicaSense RedEdge-MX Dual system and a DJI Matrice M350 RTK with a Zenmuse L2 LiDAR sensor. To assess the impact of spatial resolution on disease identification accuracy, multispectral surveys were conducted at altitudes of 40, 57, and 72 m AGL, resulting in GSDs of 3, 4, and 5 cm/pixel, respectively. Surveys were conducted in June 2025 to establish baseline pre-symptomatic conditions and repeated in September 2025 during the pre-harvest period, when symptoms were clearly visible. LiDAR data was collected once to characterize stable structural parameters, such as tree height and crown complexity.

For tree-level analysis, precise individual tree crown delineation is essential. While point cloud-based segmentation was evaluated, a more robust delineation was achieved by integrating Deep Learning algorithms applied to RGB orthophotos. The 10-band spectral data facilitated the calculation of sensitive narrow-band indices (e.g., PRI, NDRE, Cl_RE) to detect changes in pigmentation and photosynthetic efficiency. Finally, the study applies multivariate statistical analysis to cluster trees by fusing 3D structural metrics derived from LiDAR with spectral indices. This approach aims to model species-specific stress responses and categorize cultivars based on their physiological and structural characteristics, providing a foundation for improved precision agriculture workflows.

---

Issa Loghmanieh is funded by the Stipendium Hungaricum scholarship under the joint executive program between Hungary and Iran.