Dataset for orange fruit detection from UAV in citrus orchards.

Dataset for orange fruit detection from UAV in citrus orchards - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Data Brief . 2026 Mar 28;66:112733. doi: 10.1016/j.dib.2026.112733 Search in PMC Search in PubMed View in NLM Catalog Add to search Dataset for orange fruit detection from UAV in citrus orchards Guillem Montalban-Faet Guillem Montalban-Faet 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Guillem Montalban-Faet 1 , Enrique Navarro-Modesto Enrique Navarro-Modesto 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Enrique Navarro-Modesto 1 , Andoni Salcedo-Navarro Andoni Salcedo-Navarro 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Andoni Salcedo-Navarro 1 , Rafael Fayos-Jordan Rafael Fayos-Jordan 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Rafael Fayos-Jordan 1 , Pablo Benlloch-Caballero Pablo Benlloch-Caballero 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Pablo Benlloch-Caballero 1 , Miguel Garcia-Pineda Miguel Garcia-Pineda 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Miguel Garcia-Pineda 1, ⁎ , Jaume Segura-Garcia Jaume Segura-Garcia 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain Find articles by Jaume Segura-Garcia 1, ⁎ Author information Article notes Copyright and License information 1 Computer Science Department, ETSE-UV, Universitat de València, València, Spain ⁎ Corresponding authors. [email protected] [email protected] Received 2025 Dec 14; Revised 2026 Mar 24; Accepted 2026 Mar 25; Collection date 2026 Jun. © 2026 The Author(s) This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). PMC Copyright notice PMCID: PMC13090671  PMID: 42007090 Abstract Accurate fruit detection in citrus orchards is essential for yield estimation, precision harvesting, and automated orchard monitoring. Although UAV-based imaging has become a powerful tool in precision agriculture, publicly available datasets for orange fruit detection remain scarce, particularly those integrating multispectral data under real field conditions. This lack of open resources limits the development and benchmarking of robust deep-learning models for cross-spectral and illumination-invariant detection. We present CampanetaOrangeFruit, a dataset acquired with a DJI Mavic 3 Multispectral UAV flying at 14 m above ground level over a commercial citrus orchard in Corbera, Valencia, Spain. The dataset comprises 550 synchronized captures (RGB + four multispectral bands: R, G, RE, NIR) for a total of 2750 images and 301,232 annotated orange instances. Each image includes YOLOv5-format annotations generated through a homography-based reprojection process, ensuring geometric consistency across spectral modalities. CampanetaOrangeFruit uniquely provides pixel-aligned, cross-spectral UAV imagery with fine-grained fruit-level annotations, enabling research on fruit detection, yield estimation, and domain adaptation in real-world orchard environments. It represents a valuable benchmark for advancing deep-learning approaches in precision agriculture and sustainable citrus production. Keywords: Multispectral imaging, Orange fruit, Precision agriculture, Citrus orchard, UAV Specifications Table Subject Computer Science Specific subject area Citrus Detection, Precision Agriculture, Remote Sensing, Photogrammetry Type of data Image files (raw JPG/TIFF), annotation text files, scripts and figures Data collection • Platform: DJI Mavic 3 Multispectral UAV • Altitude: 14 m Above Ground Level • Speed: 1.2 m/s • Overlap: 80% frontal, 70% lateral • Ground Sampling Distance: 0.65 cm/pixel Data source location Corbera, Valencia, Spain {39.176902, −0.353776; 39.176283, −0.351407} Data accessibility 10.5281/zenodo.17866343 under CC-BY-4.0 licence Related research article The associated research article with this dataset is under preparation Open in a new tab 1. Value of the Data • The Campaneta-Orange-Fruit dataset provides a large-scale UAV-based collection of RGB and multispectral imagery of orange orchards acquired under real field conditions, offering a unique aerial perspective that enables detailed fruit-level analysis rarely available in open-access agricultural datasets. • It comprises 2750 pixel-aligned images across five spectral bands (RGB, G, R, RE, NIR) and over 301,000 annotated oranges, representing one of the most extensive public resources for citrus fruit detection, yield estimation, and precision horticulture research. he repository has an approximate size of 21 GB. • The multispectral configuration supports the study of cross-spectral modeling, illumination invariance, and canopy–fruit reflectance interactions, expanding the potential of deep-learning models to generalize under diverse spectral and lighting conditions. • Instance-level YOLO annotations and accompanying metadata, flight logs, calibration parameters, and Python scripts for homography estimation and label reprojection ensure geometric consistency across modalities and promote full reproducibility in cross-spectral experiments. • Potential downstream applications include automated fruit counting, yield mapping, and precision harvesting, advancing sustainable orchard management and UAV-based monitoring for intelligent, data-driven agriculture. • The dataset supports practical yield estimation and production quantification for citrus growers by enabling automated fruit detection and counting at tree and orchard scale. Such UAV-based production assessment can assist farmers in harvest scheduling, labor and logistics planning, and market forecasting, contributing to data-driven decision-making and improved economic sustainability in commercial orchards. 2. Background Precision agriculture increasingly leverages remote sensing to optimize orchard management, enabling non-destructive monitoring of fruit load, canopy health, and harvest timing in a cost-efficient and scalable way [ 1 ]. Recent advances in digital agriculture technologies—including 5G-IoT communication frameworks, intelligent sensor networks, and interactive visualization platforms—are further transforming precision farming by enabling real-time data acquisition, secure connectivity, and scalable analytics for decision support [ [2] , [3] , [4] ]. These innovations integrate UAV imagery, ground sensors, and cloud infrastructures into unified monitoring systems, enhancing automation, interoperability, and responsiveness across agricultural environments. Unmanned aerial vehicles (UAVs) equipped with multispectral or RGB sensors have become key tools in orchard-scale monitoring because they deliver high spatial resolution and flexible deployment compared with satellite or ground-based methods [ 5 , 6 ]. In citrus orchards, UAV-based imagery has been successfully used for fruit detection, fruit counting, and yield estimation, providing accurate, non-invasive measures of production potential [ [7] , [8] , [9] , [10] , [11] ]. Datasets such as CitDet, YOLOC-tiny, and YOLOv8-Scm have enabled model benchmarking for fruit counting, ripeness recognition, and even sunburn detection under field conditions [ [12] , [13] , [14] ]. Direct fruit-detection approaches from aerial imagery support pre-harvest yield prediction and facilitate decision-support systems in commercial orchards [ 9 , 11 , 12 ]. Multispectral imaging — capturing not only RGB but also red-edge and near-infrared bands — provides additional spectral information that helps discriminate fruits, canopy, and background, improving robustness to illumination changes and occlusions [ 15 , 16 ]. Integrating these narrow-band cues with vegetation indices or 3-D canopy structure enhances feature extraction and overall detection accuracy [ 13 , 17 ]. Recent deep-learning frameworks, including improved YOLOv7–v9 models, convolutional networks, and attention-based architectures, have demonstrated excellent performance for orchard imagery analysis, enabling the detection of individual fruits and estimation of yield at tree or orchard level [ 7 , 9 , 11 , 14 , 21 ]. These methods are further strengthened through multisensor fusion and IoT-based UAV networks, which allow real-time diagnosis and disease detection in citrus orchards [ 18 , 19 ]. Despite these advances, there remains a shortage of publicly available, high-resolution UAV datasets for citrus fruit detection that combine multispectral data, multi-temporal acquisitions, and instance-level annotations. Existing repositories, such as CampanetaWeed [ 20 ]. or EscaYard [ 6 ], focus on weed or disease detection but rarely include fruit-specific labeling or pixel alignment. Other datasets, such as CitDet [ 12 ] or SootyMold [ 15 ], provide valuable RGB data but lack multi-band spectral coverage. To fill this gap, the CampanetaOrangeFruit dataset provides pixel-aligned RGB + multispectral (R, G, RE, NIR) imagery of citrus orchards captured over UAV flight. The dataset includes instance-level annotations of individual oranges, enabling robust research on fruit detection, yield estimation, temporal generalization, and spectral-domain adaptation in real-world orchard settings. Table 1 contrasts CampanetaOrangeFruit with other publicly available citrus datasets, emphasizing its unique combination of drone-based pixel-aligned multispectral imagery, multi-temporal coverage, and instance-level annotations of individual oranges. While existing collections such as CitDet, YOLOC-tiny, YOLOv8-Scm, and SootyMold rely primarily on ground-level RGB imagery captured under single-date or controlled conditions, CampanetaOrangeFruit provides five-band (RGB + R, G, RE, NIR) UAV imagery from three separate flight dates. This configuration enables cross-spectral analysis, temporal generalization, and robust benchmarking of fruit-detection and yield-estimation models in real orchard environments. Table 1. Comparison of datasets. Dataset Name Year Crop Type Spectral Bands Num. of images Num. of annotations Labeling Scheme Target Classes Notes Campaneta OrangeFruit (this work) 2025 Citrus orchard (orange) RGB + R, G, RE, NIR 2 750 301 232 YOLOv5 Orange UAV based, pixel-aligned, cross spectral imagery CitDet 2023 Citrus (HLB-affected orchards) RGB 579 Over 32 000 COCO / YOLO Orange Ground-based RGB imagery (on-tree and fallen citrus fruits) for fruit detection benchmarking YOLOC-tiny Citrus Dataset 2024 Citrus (navel orange, Ehime Jelly orange, Harumi tangerine) RGB 3 025 10 653 YOLOv7-tiny Citrus at multiple ripeness levels Ground-based RGB imagery of multiripeness citrus fruits YOLOv8-Scm Citrus Dataset 2025 Citrus (sunburn detection) RGB 2 099 5 390 YOLOv8 Healthy vs. sunburned oranges Ground-based RGB imagery for classification of healthy vs. sunburned citrus fruits SootyMold Citrus Canopy Dataset 2023 Citrus canopy RGB (night-vision) 1 000 4 500 YOLOv5 / YOLOv7 Healthy vs sooty-mold-infected leaves Ground-based RGB dataset for detecting sooty mold on citrus canopies Open in a new tab 3. Data Description 3.1. Study area Fig. 1 shows the geographic extent of the citrus orchard in Corbera, Valencia, Spain, where multispectral flights were conducted. Fig. 1. Open in a new tab Study area map illustrating the orchard boundary and UAV flight coverage in Corbera, Valencia (latitude 39.1769 N, longitude 0.3538 W). The region is characterized by a Mediterranean climate, with mild winters and hot, dry summers, and rainfall concentrated mainly in autumn. Reported long-term averages for the nearby coastal Valencia area indicate an annual mean temperature of ∼17–19 °C and annual precipitation of ∼430–455 mm, depending on the reference station/dataset. 3.2. Repository structure In Fig. 2 , we present the complete directory structure of the Campaneta-Orange-Fruit dataset. The root directory contains five main subfolders — RGB/, G/, R/, RE/, and NIR/ — corresponding to the spectral bands captured during the UAV flight. Each folder includes the image files (.JPG for RGB and .TIF for multispectral bands), the associated YOLO-format annotation files (.txt), and a classes.txt file defining the object categories [ 19 ]. Fig. 2. Open in a new tab Repository structure of directories and files of the dataset. Additionally, the figures/ folder includes two illustrative images — Figure_1.png showing the study area and UAV flight coverage in Corbera (Valencia), and Figure_2.png summarizing the directory structure of annotated instances across spectral bands — while the scripts/ directory provides two Python utilities ( estimate_band_homographies.py and transfer_yolo_labels.py ) for computing inter-band homographies and transferring YOLO annotations, ensuring accurate pixel-level alignment across all multispectral modalities. 3.3. Class distribution Table 2 summarises the number of annotated instances and statistics for class Orange. Table 2. Distribution of annotated objects per spectral band. Band Class name Instances (labels) Number of images Mean labels per image Minimum labels per image Maximum labels per image RGB Orange 80,446 550 146.27 1 776 G Orange 55,384 550 100.70 1 714 NIR Orange 54,912 550 99.84 1 714 RE Orange 55,158 550 100.29 1 713 R Orange 55,332 550 100.60 1 714 Open in a new tab 4. Experimental Design, Materials and Methods 4.1. Dataset construction workflow The data acquisition process began with the definition of the target area, which was incorporated into a pre-planned UAV mission to ensure systematic coverage of the orange orchard. The DJI Mavic 3 Multispectral was configured to capture both RGB and multispectral imagery across four narrow bands — Green (G), Red (R), Red-Edge (RE), and Near-Infrared (NIR) — providing complementary spectral information for analysis. Following the UAV flight, all collected images were screened for quality and content, retaining only those where oranges were clearly visible and properly exposed. The selected subset was then annotated using LabelImg, an open-source labeling tool. The annotation team performed a detailed review of each RGB image, zooming in to accurately delineate visible oranges. These annotations were then transferred to the multispectral bands using the provided reprojection scripts, ensuring geometric consistency across all modalities. Fig. 3 illustrates the structured workflow applied during the data acquisition and labeling stages, from mission planning to expert annotation review, culminating in the reprojection of annotations onto the multispectral bands to ensure cross-spectral consistency. Fig. 3. Open in a new tab Dataset construction workflow followed, from mission planning to multispectral label reprojection. 4.1.1. UAV flight and data acquisition The image acquisition campaign was conducted under stable and cloudless conditions at midday to minimize shadows and illumination variability. According to this, the DJI Mavic 3 Multispectral mission was executed using a 2D single-grid flight plan designed for nadir imaging and uniform orchard coverage. A double-grid configuration was not required, as the primary objective was high-resolution fruit detection rather than 3D canopy modeling. The camera was oriented in a nadir position (gimbal angle −90°) throughout the mission to minimize geometric distortion and ensure consistent top-down perspective across all captures. The UAV flew at a constant altitude of 14 m above ground level with 80% frontal and 70% lateral overlap, at a ground speed of approximately 1.2 m/s. The total flight time for the mapped area was approximately 40 min (including takeoff and landing), resulting in 550 synchronized RGB and multispectral captures under stable midday illumination conditions. The onboard imaging system integrates five synchronized 4/3″ CMOS sensors: one 20-MP RGB camera and four monochromatic 5-MP sensors centred at 560 nm (G), 650 nm (R), 730 nm (RE), and 860 nm (NIR), each with a 10-nm full-width half-maximum (FWHM). The resulting ground sampling distance (GSD) was approximately 0.65 cm/pixel, allowing individual fruit detection from UAV height. A total of 550 captures were selected from the mission, each containing one RGB frame and four multispectral bands synchronized in time and geometry. 4.2. Radiometric calibration The DJI Mavic 3 Multispectral system provides radiometrically calibrated imagery through its integrated sunlight sensor and factory calibration parameters. During acquisition, the onboard upward-facing irradiance sensor continuously measured incident solar radiation, enabling per-capture compensation for illumination variability. The multispectral bands (G, R, RE, NIR) were stored as 16-bit TIFF images containing radiometrically corrected reflectance values proportional to surface reflectance. Standard manufacturer-provided calibration metadata (including gain, exposure time, and irradiance normalization factors embedded in the image headers) were preserved without additional post-processing. No empirical reflectance panel calibration was applied, as flights were conducted under stable, cloudless midday conditions to minimize illumination changes. Therefore, the dataset provides radiometrically consistent band measurements suitable for vegetation index computation, cross-spectral modeling, and reflectance-based feature extraction, while users requiring absolute reflectance validation across campaigns may optionally perform additional panel-based normalization. 4.3. Annotation methodology Fruit labeling was performed using LabelImg, an open-source tool compatible with the YOLO annotation format. The labeling procedure followed a three-stage verification protocol to ensure consistency and accuracy: 1. Four trained annotators independently labeled each RGB image, identifying every visible orange. 2. Two additional reviewers then cross-checked the annotations, resolving discrepancies and generating the final consensus version. 3. Bounding boxes were saved in YOLOv5 format, producing one *.txt file per image. Each annotation file contains one line per object with five normalized values: <class> 〈x_center〉 〈y_center〉 〈width〉 〈height〉 Where: • <class_id> = 0 (Orange) • < x _center>, 〈y_center〉 = normalized bounding box center coordinates in the range [0, 1] • <width>, 〈height〉 = normalized bounding box width and height in the range [0, 1] The RGB annotations served as the reference geometry for the multispectral bands. Using the provided Python scripts, bounding boxes were later reprojected onto the G, R, RE, and NIR images via the homographies estimated between sensors, ensuring pixel-level consistency across all spectral modalities. As seen in Fig. 4 , the dataset includes UAV images with varying densities of annotated oranges, reflecting the natural variability of fruit distribution across the orchard. Panels (a–c) illustrate canopies ranging from sparsely to densely labeled regions, which demonstrate differences in tree structure, canopy coverage, and fruit clustering captured during the UAV flights. Fig. 4. Open in a new tab Pictures with different density of labels. As shown in Fig. 5 , the dataset contains orange instances with noticeable variation in size, illumination, and occlusion. Panels (a–c) display examples where fruits appear under different lighting conditions, degrees of shadowing, and levels of partial obstruction by leaves or branches, highlighting the dataset’s visual diversity and its suitability for testing model robustness. Fig. 5. Open in a new tab Pictures with different examples of instances. 4.4. Label reprojection across spectral bands To ensure geometric consistency across the RGB and multispectral layers, a homography-based reprojection process was implemented. Homography estimation. For each RGB–band pair, keypoint correspondences were detected using the ORB algorithm, followed by RANSAC filtering to remove outliers. The resulting 3 × 3 homography matrix captures the geometric relationship between the RGB camera and each multispectral sensor, accounting for lens distortion, field-of-view differences, and gimbal offsets. To ensure full reproducibility of the homography estimation process, we report the exact configuration used in all experiments. For each RGB-to-band pair (G, R, RE, NIR), images were first converted to grayscale and processed at a downscale factor of 0.5. Keypoint correspondences were extracted with ORB using a feature limit of 4000 keypoints per image. Binary descriptors were matched with a brute-force matcher using Hamming distance and cross-check enabled. Homographies were then estimated with cv2.findHomography using the RANSAC method and a reprojection threshold of 3.0 pixels. A minimum of 30 raw matches was required before 232 running RANSAC; pairs below this threshold were discarded. The same parameterization was applied consistently across all RGB→(G, R, RE, NIR) transfers. Bounding box reprojection. Each YOLO annotation ⟨ X a , Y a , w, h ⟩ from the RGB image was converted into pixel-space corner coordinates. These corner points were then warped using the corresponding homography matrix to reproject the bounding box onto the target spectral band. The resulting quadrilateral was clipped to the band’s image extent, and boxes that fell partially or completely outside the frame, lost a substantial portion of their visible area, or became smaller than a minimum pixel threshold were excluded. Therefore, after warping the RGB bounding-box corners via the estimated homography matrices, explicit filtering criteria were applied to guarantee geometric validity and practical detectability. A reprojected box was discarded if any of the following conditions were met: (i) the warped box became degenerate (zero or negative area); (ii) the warped box lay completely outside the target image after boundary clipping; (iii) <25% of the warped box area remained visible inside the target image; or (iv) the clipped axis-aligned bounding box had a width or height smaller than 2 pixels. These thresholds were applied uniformly across all spectral bands (G, R, RE, NIR) to ensure consistency and are implemented in the label-transfer procedure. Normalization and export. Remaining boxes were renormalized with respect to the spatial dimensions of the spectral band and exported back in standard YOLO format, producing one annotation file (.txt) per multispectral image. This procedure guarantees that all spectral layers share consistent object geometry while preserving the actual parallax and optical differences between sensors. To facilitate reuse and integration into benchmarking pipelines, a structured metadata manifest is provided in the Zenodo repository. The manifest includes, for each spectral band, image resolution, bit depth, file format, number of images, and annotation statistics. RGB images have a spatial resolution of 5280 × 3956 pixels (8-bit JPG), while the multispectral bands (G, R, RE, NIR) have a resolution of 2592 × 1944 pixels (16-bit TIFF). Each band contains 550 images, and per-band annotation counts and summary statistics are reported in Table 2 . The manifest also mirrors the directory structure illustrated in Fig. 2 , enabling automated dataset parsing. Fig. 6 illustrates how the homography-driven label transfer behaves on a representative frame. The RGB image retains all original YOLO boxes, and the red overlays immediately reveal which detections disappear once warped into the G band. The adjacent G-band panel displays only the successfully mapped boxes (lime), demonstrating both the spatial consistency achieved by the transfer and the specific instances where low overlap or occlusions cause a box to be discarded. This visualization provides a qualitative sanity check that complements the quantitative band-level statistics reported earlier. Fig. 6. Open in a new tab RGB panel (left) shows all annotated oranges, with red boxes marking detections that fail the RGB G homography transfer; the right panel overlays the successfully reprojected boxes on the G-band radiance image, confirming where each detection remains valid. Finally, in order to promote fair and consistent benchmarking, we propose a reference evaluation protocol for fruit detection and yield-estimation research using CampanetaOrangeFruit. As the dataset comprises imagery acquired on three distinct flight dates, we recommend a date-based split as the primary experimental setup: training on two flight dates and testing on the remaining date. This configuration encourages temporal generalization and reduces spatial leakage between overlapping captures. As an alternative, a 70%/15%/15% random split at the image level may be used for within date benchmarking. For detection performance, we recommend reporting 275 [email protected], [email protected]:0.95, Precision, Recall, and F1-score. For yield-related tasks, Mean Absolute Error (MAE) of fruit counts per image and the coefficient of determination (R²) between predicted and ground-truth counts are suggested. These metrics align with current object-detection and precision-agriculture evaluation standards. Limitations The main limitations of the CampanetaOrangeFruit dataset are associated with the specific area and single UAV flight used for data acquisition. The imagery was collected over a single commercial citrus orchard located in Corbera (Valencia, Spain), which may limit the generalization of models to orchards with different canopy architectures, management practices, or environmental conditions. As the dataset represents a single acquisition campaign under specific illumination and phenological conditions, care should be taken when applying trained models to other orchards or seasons. From a computational perspective, scaling this methodology to substantially larger orchards would increase both data volume and processing complexity. Given the native spatial resolution of the RGB images (5280 × 3956 pixels) and multispectral bands (2592 × 1944 pixels), a large-scale campaign covering several tens of hectares could easily produce several thousands of synchronized captures, resulting in tens to hundreds of gigabytes of raw imagery. This would proportionally increase storage requirements, I/O operations, and backup demands. Photogrammetric reconstruction complexity grows superlinearly with the number of overlapping images. Structure-from-motion (SfM) alignment requires pairwise feature matching across overlapping frames, and bundle adjustment optimization scales with the number of cameras and tie points. For large datasets, memory usage and processing time may increase significantly, potentially requiring high-RAM workstations or distributed/cloud-based processing. Dense multi-view stereo reconstruction for orthomosaic generation further amplifies GPU and CPU requirements. Although homography estimation between RGB and multispectral bands is computed per sensor pair and not globally across the orchard, processing thousands of images still increases total keypoint detection, descriptor matching, and RANSAC computations. While each operation is relatively lightweight, cumulative runtime becomes non-negligible at scale. Parallelized or batched processing would therefore be recommended for large deployments. A further scalability constraint arises from manual annotation. Even with automated cross-spectral label transfer, the initial RGB labeling effort increases linearly with dataset size. Large-scale replication may therefore require semi-automated annotation strategies (e.g., model-assisted labeling, active learning, or pre-trained detector bootstrapping) to remain operationally feasible. Consequently, while the proposed methodology is technically replicable in larger orchards, efficient scaling may require high-performance hardware (multi-core CPUs, ≥64–128 GB RAM for large SfM projects, GPU acceleration for dense reconstruction), tiled or block-wise photogrammetric processing, and automated data-management pipelines to ensure computational tractability. An additional limitation is that all data were acquired at a single flight 312 altitude (14 m AGL), yielding a ground sampling distance (GSD) of approximately 0.65 cm/pixel. Multi-altitude experiments were not conducted. At higher altitudes (e.g., 50 m AGL), the expected GSD would increase to approximately 1.4 cm/pixel. While mature oranges (typically 6–8 cm in diameter) would still occupy several pixels and remain theoretically detectable, reduced spatial resolution may impair the detection of smaller or partially occluded fruits and increase confusion with surrounding foliage, particularly under complex illumination conditions. Therefore, detection robustness at higher altitudes remains an open evaluation scenario and may require model retraining or architecture adjustments. Future work could systematically assess the trade-off between spatial resolution, coverage efficiency, and detection accuracy. Ethics Statement The work involves no human participants, no live animal experimentation, and no data collected from social media platforms. All UAV operations complied with EU Regulation 2019/947 and Spanish Royal Decree 1036/2017 on remotely piloted aircraft, including flight permissions over private agricultural land and maintenance of safe distances from inhabited areas. CRediT Author Statement Guillem Montalban-Faet : Conceptualization, Methodology, Data Curation, Writing - Original Draft. Enrique Navarro-Modesto : Data curation, Writing- review & editing. Andoni Salcedo-Navarro : Data curation, Writing- review & editing. Rafael Fayos-Jordan : Data curation, Writing- review & editing. Pablo Benlloch-Caballero : Data curation, Writing- review & editing. Miguel Garcia-Pineda : Conceptualization, Data Curation, Writing - review & editing, Supervision, Funding acquisition, Project administration. Jaume Segura-Garcia : Methodology, Data Curation, Writing - review & editing, Supervision, Funding acquisition, Project administration. Declaration of AI Use During the preparation of this work, the authors used Open AI’s ChatGPT (GPT-5.0) to summarize and edit content. After using this tool, the authors reviewed and edited the content as needed and they take full responsibility for the content of the publication. Acknowledgements This research has been partially supported by the Spanish Ministry of Science and Innovation/Spanish Research Agency (MCIN/AEI) within the project Agriculture 6.0 with reference TED2021-131040B-C33, funded by MCIN/AEI/ 10.13039/501100011033 and by the European Union “NextGenerationEU”/PRTR. Also, it has been supported by the project DRONIA with reference INREIA/2024/164 funded by Generalitat Valenciana in the RETECH IA programme, with funds from AEI and European Union NextGenerationEU/PRTR, and the research stay in a company CIAEST/2024/110, funded by Generalitat Valenciana. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Contributor Information Miguel Garcia-Pineda, Email: [email protected]. Jaume Segura-Garcia, Email: [email protected]. Data Availability Zenodo Campaneta-Orange-Fruit (Original data) . References 1. Ambaru M., Manvitha M., et al. Synergistic integration of remote sensing and soil metagenomics data: advancing precision agriculture through interdisciplinary approaches. Front. Sustain. Food Syst. 2025 [ Google Scholar ] 2. Fayos-Jordan R., Araiz-Chapa R., Felici-Castell S., Segura-Garcia J., Perez-Solano J.J., Alcaraz-Calero J.M. ECO4RUPA: 5G-IoT inclusive and intelligent routing ecosystem with low-cost air quality monitoring. Information, 2023;14(8):445. doi: 10.3390/info14080445. [ DOI ] [ Google Scholar ] 3. Benlloch-Caballero P., Sanchez-Navarro I., Matencio-Escolar A., Calero J.M.A., Wang Q. Topology-Aware cognitive self-protection framework for automated detection and mitigation of security and privacy incidents in 5G-IoT networks. 2023 IEEE 31st International Conference on Network Protocols (ICNP); Reykjavik, Iceland; 2023. pp. 1–6. [ DOI ] [ Google Scholar ] 4. Salcedo-Navarro A., Montalban-Faet G., Segura-Garcia J., Garcia-Pineda M. A scalable web system for multi-index 3D point cloud visualization and real-time sensor monitoring in precision agriculture. SoftwareX. 2025;32 doi: 10.1016/j.softx.2025.102443. [ DOI ] [ Google Scholar ] 5. Grbović Ivošević M., et al. Integrating UAV multispectral imaging and proximal sensing for high-precision crop monitoring. PLoS. One. 2025;20 doi: 10.1371/journal.pone.0322712. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Vélez S., Ariza-Sentís M., Valente J. EscaYard: precision viticulture multimodal dataset of vineyards affected by Esca disease. Data Brief. 2024;54 doi: 10.1016/j.dib.2024.110497. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Daiaeddine M.J., Badrouss S., El Harti A., Bachaoui E., Biniz M., Mouncif H. UAV imagery, advanced deep learning, and YOLOv7 object detection model in enhancing citrus yield estimation. Foods Raw Mater. 2024 [ Google Scholar ] 8. Zhai J., Jing Y. Deriving early citrus fruit yield estimation by combining multiple growing-period data and improved YOLOv8 modeling. Sensors. (Basel) 2025;25:4718. doi: 10.3390/s25154718. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Kwon S.-H., Ku K., Le A.T., et al. Enhancing citrus fruit yield investigations through flight-height optimization with UAV imaging. Sci. Rep. 2024;14 doi: 10.1038/s41598-023-50921-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Wang L., Lu J., et al. In situ nondestructive identification of citrus fruit ripeness via hyperspectral imaging technology. Plant Methods. 2025;21:1354. doi: 10.1186/s13007-025-01354-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Osco L., Nogueira K., et al. Semantic segmentation of citrus-orchard using deep neural networks and multispectral UAV-based imagery. Precis. Agric. 2022;23:1046–1062. [ Google Scholar ] 12. James J.A., Manching H.K., Mattia M.R., Bowman K.D., Hulse-Kemp A.M., Beksi W.J. CitDet: A benchmark dataset for citrus fruit detection. 2023 arXiv:2309.05645. [ Google Scholar ] 13. Tang Z., Xu L., Li H., Chen M., Shi X., Zhou L., Wang Y., Wu Z., Zhao Y., He Y., Ma W., Yang N., Luo L., Qiu Y. YOLOC-tiny: a generalized lightweight real-time detection model for multiripeness fruits of large non-green-ripe citrus. Front. Plant Sci. 2024;15 doi: 10.3389/fpls.2024.1415006. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Cong Y., Chen X., et al. YOLOv8-Scm: an improved model for citrus fruit sunburn identification and classification in complex natural scenes. Front. Plant Sci. 2025;16 doi: 10.3389/fpls.2025.1591989. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Apacionado B.V., Ahamed T. Sooty mold detection on citrus tree canopy using deep learning algorithms. Sensors. 2023;23:8519. doi: 10.3390/s23208519. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Zhang W., Wang J., Liu Y., et al. Deep-learning-based in-field citrus fruit detection and tracking. Hortic. Res. 2022;9 doi: 10.1093/hr/uhac003. uhac003. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Saif R., Chancia C., et al. Exploring UAS imaging modalities for precision agriculture: predicting root yield and disease severity using multispectral, hyperspectral, and LiDAR sensing. Algorithms, Technol. App. Multispectral Hyperspectral Imag. XXXI. 2025 [ Google Scholar ] 18. Foroughi A., Lloret J., Jimenez J., Sendra S. An edge computing wireless sensor network for diagnosing orange fruit disease. Cluster. Comput. 2025:28. doi: 10.1007/s10586-024-04999-y. [ DOI ] [ Google Scholar ] 19. Salcedo-Navarro A., Montalban-Faet G., Garcia-Pineda M., Segura-Garcia J. Dataset for weed detection in fruit orchards. Data Brief. 2025;63 doi: 10.1016/j.dib.2025.112276. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Bascon M.V., Nakata T., Shibata S. Estimating yield-related traits using UAV-derived multispectral images to improve grain yield prediction. Agriculture. 2022;12:1141. [ Google Scholar ] 21. Wang H., Hou D., et al. Estimation of maize yield in plastic-film-mulched field using UAV multispectral imagery. Agronomy. 2025;15 [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Zenodo Campaneta-Orange-Fruit (Original data) . Articles from Data in Brief are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (2.0 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top